Projection-based optimization offers a gradient-free alternative to conventional gradient-based methods, but existing frameworks suffer from high computational costs and poor scaling with depth. We introduce $\mathcal{P}$Torch, a projection-based framework that builds directly on top of PyTorch's autograd engine, eliminating the overhead of prior approaches. We further improve performance through nonlinear relaxation and memory-efficient projections. Together, these contributions make the algorithm competitive with gradient-based methods for shallow MLPs on MNIST and CIFAR-10. Finally, we prove a vanishing target theorem showing that the layer-wise learning signal decays with increased depth, serving as an analogue of the vanishing gradients problem for projection-based methods.

Given the profound impacts on human health and ecosystems, air pollution necessitates robust quality prediction to inform policy-making. Traditional physics-based forecasting relies on idealized closed-system assumptions, whereas data-driven approaches often lack interpretability. Recent hybrid methods, however, neglect data homogeneity, resulting in redundancy, parameter inefficiency, and gradient conflicts. Moreover, high-concentration regions inherently exhibit drastic fluctuations, making it difficult to explicitly distinguish them from the overall pollution field, especially when using spatial information alone. To address these limitations, a Physics-Anchored Spatial-Spectral ($PAS^2$) paradigm is proposed. Specifically, a collaboration scheme is firstly elaborated, within which a physical estimator captures deterministic trends, anchoring the neural branch’s input space and directing it to learn complementary fluctuations. Building on this, we equip the neural branch with a Spatial–Spectral ($S^2$) Reasoner that integrates graph operators to model local spatial diffusion and spectral operators to explicitly capture the drastic fluctuations in high-concentration regions. Experiments on real-world datasets show that $PAS^2$ consistently improves over strong data-driven and physics-guided baselines, especially under volatile sudden-change scenarios. Codes are attached and will be released at GitHub.

Generating property-optimized mRNA sequences is central to applications such as vaccine design and protein replacement therapy, but remains challenging due to limited data, complex sequence-function relationships, and the narrow space of biologically viable sequences. Generative methods that drift away from the data manifold can yield sequences that fail to fold, translate poorly, or are otherwise nonfunctional. We present RNAGenScape, a property-guided manifold Langevin dynamics framework for mRNA sequence generation that operates directly on a learned manifold of real data. By performing iterative local optimization constrained to this manifold, RNAGenScape preserves biological viability, accesses reliable guidance, and avoids excursions into nonfunctional regions of the ambient sequence space. The framework integrates three components: (1) an autoencoder jointly trained with a property predictor to learn a property-organized latent manifold, (2) a denoising autoencoder that projects updates back onto the manifold, and (3) a property-guided Langevin dynamics procedure that performs optimization along the manifold. Across three real-world mRNA datasets spanning two orders of magnitude in size, RNAGenScape increases median property gain by up to 148% and success rate by up to 30% while ensuring biological viability of generated sequences, and achieves competitive inference efficiency relative to existing generative approaches.

Hartigan's $k$-means algorithm has theoretical advantages over the standard Lloyd's variant, which has become synonymous with ``$k$-means''. Yet, Hartigan's algorithm is far slower in practice due to its inherently sequential calculations. We devise a Batched Hartigan's method that enables vectorized calculation by exploiting the triangle inequality to expose a set of parallel-safe updates and a smaller set of sequential updates. This makes Hartigan's method competitive with Lloyd's in execution speed for the first time.

Long-form omni-modal video understanding requires models to integrate vision, speech, and ambient audio with coherent long-context reasoning. Existing video benchmarks often trade off temporal scale, modality coverage, open-ended interaction, and interpretable scoring. To address this gap, we introduce $\textbf{LongShOTBench}$, a long video evaluation benchmark designed around three coupled goals: holistic omni-modal integration, intent-driven open-ended interaction, and rubric-level diagnosis. It constructs single- and multi-turn questions from practical viewing scenarios, involving systematical tasks for probing visual, speech, ambient-audio, temporal, and cross-modal reasoning. Each item includes a reference answer and a weighted criterion-level rubric, enabling evaluation to identify which perceptual facts, temporal links, modality-grounding requirements, and reasoning steps are satisfied or missed. All samples are manually verified and corrected to improve grounding, clarity, and rubric reliability. We also introduce $\textbf{LongShOTAgent}$, a training-free omni-modal evidence-seeking agent that couples full-video preprocessing with targeted retrieval, query-adaptive segment refinement, and explicit claim verification over visual, speech, and non-speech audio evidence. Its iterative search-refine-verify loop exposes intermediate evidence and lets modality-specific specialists re-analyze relevant moments before answering. We perform comprehensive evaluation of 105 video-capable models spanning open-source omni-modal models, vision-language systems, audio LLMs, agentic pipelines and closed-source APIs. Across this broad evaluation, current MLLMs remain far from saturating LongShOTBench, while our LongShOTAgent emerges as the strongest training-free system, reaching 66.64\% overall. By releasing the benchmark, leaderboard, and agentic method, our work provides the community with a shared, interpretable testbed for evaluating and advancing long-form omni-modal video reasoning.

Current LLM-for-GPU benchmarks largely measure CUDA fluency on isolated operators or short kernel-generation tasks, whereas practical GPU performance engineering often starts from trusted sequential CPU programs and requires correct, end-to-end acceleration. To fill this gap, we introduce \textsc{AccelEval}, a benchmark for whole-program CPU-to-CUDA acceleration. AccelEval contains 42 tasks derived from public repositories and established HPC/application suites across six domains, each equipped with deterministic correctness oracles and three input scales. Beyond pass/fail correctness, AccelEval reports end-to-end speedup over CPU baselines and a novel peak-attainment metric which normalizes each model’s speedup against the best correct solution. Using this framework, we evaluate eight frontier LLMs across multiple input scales, GPU architectures and sample rates. At pass@1 on NVIDIA H200 at medium scale, the strongest model passes 38/42 tasks at 71.2$\times$ geometric-mean speedup over the CPU baseline. On the subset of 11 tasks with human-written CUDA references, the best-of-8 LLM oracle reaches 0.75$\times$ of human CUDA performance. To make CUDA-optimization behavior interpretable and reusable, we decompose correct solutions into 43 CUDA acceleration patterns and show that feeding winner-derived pattern guidance back to models improves paired geometric-mean speedup by 1.78$\times$, compared with 1.18$\times$ for a generic-tips control. Together, these results establish AccelEval as a reproducible diagnostic platform for end-to-end LLM-assisted GPU acceleration. An anonymized open-source release is available at \url{https://anonymous.4open.science/r/AccelEval-2397}.

State-of-the-art neural network sparsification relies on alternating dense and sparse training phases. While computationally expensive, dense phases are critical for performance. Our empirical analysis reveals the mechanics behind this gain: upon reintroduction, previously masked parameters receive disproportionately large gradients that drive essential mask rewiring and parameter sign-flipping. Building on these insights, we optimize the performance-to-FLOPs Pareto frontier through three innovations: (1) zero-initializing masked parameters during dense phases to promote relevant sign flips, (2) truncating dense phases based on early mask convergence, and (3) replacing the dense phase entirely with a sparse rewiring phase that updates only the active mask and a small, high-gradient fraction of out-of-mask weights. Evaluated on CIFAR100 and ImageNet, our variant ARC (Alternating Rewiring Compression) matches baseline AC/DC performance at a fraction of the computational cost, establishing a superior performance-to-FLOPs frontier also compared to standard sparse-to-sparse algorithms.

Video is temporally redundant: adjacent frames usually share most objects, background, and layout. Yet existing video multimodal large language models (video MLLMs) usually encode each sampled frame as an independent RGB image, causing visual tokens to repeat content already present in earlier frames. This suggests a more direct video interface: send a full reference frame only when the scene cannot be predicted well from prior context, and otherwise transmit a compact description of inter-frame changes. We call this interface a predictive visual code, and instantiate it for video MLLMs as AdaCodec. AdaCodec spends full visual tokens on a reference frame only when its conditional predictive cost is high; otherwise, it encodes inter-frame changes, including motion and prediction residuals, as compact P-tokens. Across all eleven benchmarks, AdaCodec improves over the Qwen3-VL-8B per-frame RGB baseline at a matched visual-token budget. Even at $1/7$ the budget, AdaCodec with 32k tokens surpasses the 224k baseline on all long-video benchmarks; on five general-video benchmarks, it raises the average score while substantially cutting time-to-first-token from 9.26s to 1.62s.

Partial Label Learning (PLL) trains models from instances with candidate label sets containing the ground-truth, while Semi-Supervised Partial Label Learning (SSPLL) further leverages the available unlabeled data to prompt the performance. The previous SSPLL methods implicitly assume the unlabeled data are class-balanced distributed. However, in realistic scenarios, the class distribution of unlabeled data is unknown, resulting in performance collapse on minority classes for the existing SSPLL methods. We identify two coupled biases behind this failure: (i) pseudo-label selection bias, where high-confidence filtering tends to over-select samples from dominant classes, causing the training data to become increasingly skewed toward these classes; and (ii) disambiguation bias, where dominant classes are more likely to win within candidate sets, even when they are incorrect. Theoretically, we show that these two biases reinforce each other over training, forming a feedback loop that degrades generalization performance. To weaken this loop, we propose AdapMatch, which uses DAS to reduce the skew in training data caused by pseudo-label selection bias via prior-aware pseudo-label admission and CAD to lower the disambiguation error by suppressing candidate-set domination. Across five benchmarks, AdapMatch yields gains up to 41.31\% at most , showing strong robustness to varying scale, imbalance, and ambiguity. The code is available at \url{https://anonymous.4open.science/r/AdapMatch-8BC5}.

Graph Domain Adaptation (GDA) has emerged as an effective technique for transferring knowledge from a label-rich source graph to an unlabeled target graph by mitigating cross-graph discrepancies. However, in real-world scenarios, the attributes of partial nodes may be unobserved (i.e., missing) due to factors such as privacy protection. This inevitably exacerbates the distribution discrepancy of nodes between graph domains, making it more challenging to transfer effective knowledge to the target graph. To tackle this issue, we propose an adaptive attribute completion approach with representation space for incomplete graph domain adaptation, namely AC-GDA. It mutually assists the attribute completion and dynamically optimizes cross-graph trustworthy embeddings in an iterative manner, promoting positive knowledge migration while completing attributes. Specifically, we first propose a joint completion mechanism guided by topology similarity, which leverages cross-graph trustworthy attribute embeddings with imputation weights to adaptively complete missing node attributes. We then enhance node embeddings with label information to provide supervision for completed nodes, making nodes of the same category distribute consistently across graph domains. Furthermore, we introduce the complement confidence score to dynamically adjust the scope of trustworthy embeddings, so as to improve the quality of subsequent cross-graph imputation. Extensive experiments on a range of cross-graph tasks validate the effectiveness of AC-GDA.

Neural operators have shown strong performance in learning solution mappings for differential equations and offer fast inference across parameter spaces. However, they often suffer from spectral bias, leading to poor resolution of high-frequency and localized features, especially in out-of-distribution settings. In this work, we propose adaptive multiscale operator correction (AMOC), a hybrid framework that combines neural operator learning with physics-constrained optimization in a low-dimensional, instance-specific adaptive spectral subspace. First, a neural operator learns a coarse solution, which is then projected onto a sparse multiscale basis using either wavelet family or Fourier modes, where dominant components are selected based on energy. Finally, a physics-informed optimization is performed on adaptive spectral subspace, significantly lowering computational cost while improving accuracy. The proposed approach enables efficient correction of neural operator predictions by leveraging the sparsity and localization of multiscale representations, while bypassing automatic differentiation in the residual loss through closed-form derivatives of the chosen basis. The proposed AMOC is evaluated on the heat, Poisson, Darcy, and Helmholtz equations, achieving up to four orders-of-magnitude error reduction over FNO, with particularly strong gains observed in challenging out-of-distribution regimes.

Rigorous uncertainty quantification is essential for the safe deployment of autonomous systems in unconstrained environments. Conformal Prediction (CP) provides a distribution-free framework for this task, yet its standard formulations rely on exchangeability assumptions that are violated by the distribution shifts inherent in real-world robotics. Existing online CP methods maintain target coverage by adaptively scaling the conformal threshold, but typically employ a static nonconformity score function. We show that this fixed geometry leads to highly conservative, volume-inefficient prediction regions when environments undergo structural shifts. To address this, we propose \textbf{AdaptNC}, a framework for the joint online adaptation of both the nonconformity score parameters and the conformal threshold. AdaptNC leverages an adaptive reweighting scheme to optimize score functions, and introduces a replay buffer mechanism to mitigate the coverage instability that occurs during score transitions. We evaluate AdaptNC on diverse robotic benchmarks involving multi-agent policy changes, environmental changes and sensor degradation. Our results demonstrate that AdaptNC significantly reduces prediction region volume compared to state-of-the-art threshold-only baselines while maintaining target coverage levels.

Continual pre-training (CPT) enables injecting domain-specific knowledge into large language models during the learning rate annealing phase. However, this compute-efficient paradigm forces practitioners to navigate a challenging optimization space: simultaneously choosing the injection fraction $\lambda$, the data mixture ratio $r$, and the re-warmup peak learning rate $\eta$. Because existing scaling laws have yet to comprehensively integrate and couple these interacting variables, they cannot reliably forecast the resulting trade-offs and final convergence outcomes across unseen injection configurations. To bridge this gap, we introduce the Annealing-phase Domain Injection Scaling Law (ADIS-Law), a parametric framework that explicitly models late-stage distribution shifts as physical perturbations applied to the base pre-training trajectory. ADIS-Law formalizes two opposing macroscopic behaviors: general-domain performance follows a linear forgetting-recovery process, whereas target-domain adaptation adheres to a power-law adaptation trend. Across data-budget and model-scale extrapolation, ADIS-Law predicts target-domain trajectories with an average Spearman correlation of $0.947$ and forecasts tail losses with relative error below $5\\%$. We further show that ADIS-Law can predict optimal injection strategies at extrapolated scales: it ranks target-scale candidate policies with a mean Spearman correlation of $0.936$. By replacing blind grid search over $(\lambda, r, \eta)$ with a single fitted law, ADIS-Law offers a data-driven approach for strategy selection in annealing-phase domain injection.

Large language model (LLM)-native advertising has emerged as a viable monetization channel as LLM service providers seek sustainable revenue streams. However, integrating native ads into LLM responses introduces two largely understudied challenges: (1) controlling ad intensity---how prominently advertising content appears---and (2) ad labeling---identifying which segments of the generated text constitute advertising. We propose AdKnob, a framework jointly addressing both challenges. For ad intensity control, we introduce new ad control tokens without pretrained semantic associations and align the model via direct preference optimization to generate ads at each desired level. For ad labeling, we propose an attention rollout-based attribution mechanism that identifies ad segments by tracing their attribution to input ad-related tokens, requiring no additional inference cost. We further construct MI-KnobSet, the first multi-level ad intensity preference dataset for LLM-native ads. Experiments across six LLMs and human evaluations show that AdKnob achieves uniformly spaced ad intensities where baselines collapse to a narrow range, achieves the highest ad labeling accuracy while being over 500$\times$ faster than the strongest baseline, and is preferred over baselines by human evaluators up to 89.6\% of the time. Code and data will be publicly available.

Document QA agents do not fail only at the final answer. They can also write unsupported claims into public working memory, and those writes can affect later retrieval and synthesis. We introduce ADMIT (Auditable Decoupled Memory-write admission for Intermediate Traces), a low-budget admission controller for public OCR/text evidence packs. ADMIT applies a deterministic substring-support check $S_{\text{str}}(\hat{y}, B) > 0$ to each public commit and allows one bounded RepairRead. If repair still fails, ADMIT uses a memory-fail-closed / output-fail-open rule: it writes a $\bot$ (NO_COMMIT) sentinel to memory while the final-answer surface may still fall back to the baseline answer. On held-out M3DocVQA, a retrained final-row router closes the answer-surface gap; ADMIT's distinguishing evidence is therefore memory-surface contamination, not final-answer accuracy. The lexical invariant guarantees zero-substring accepted intermediate contamination by construction; the contribution is therefore the public memory-write boundary as a low-cost control surface with decoupled fail actions, not the invariant itself. In a 7-arm dynamic stress diagnostic on three agents, ADMIT enforces the invariant and preserves raw EM at this operating point (zero contaminated trajectories on every agent and no detectable raw-EM loss relative to the fail-open ablation), whereas final-only baselines still leave $12$--$32%$ contamination. ADMIT enforces lexical auditability, not semantic truth: relation-level entailment, citation-aware faithfulness, and visual-region grounding are not claimed, and lower-density VisDoMSlide and text-only FRAMES are reported as boundary cases.

Positional encoding in Vision Transformers presents a recurring trade-off: spatial relative-position tables (Swin) carry thousands of parameters per block and require ad-hoc interpolation to transfer across resolutions, while no-bias designs leave the attention logits geometry-agnostic. We propose a relative positional bias parameterized directly in the frequency domain. This parameterization yields three benefits over the standard spatial-table position encoding: (i) it expresses the same translation-equivariant bias with $3.25\times$ fewer parameters than Swin's spatial table at the same head count; (ii) it admits a closed-form, structurally exact transfer to a different patch grid, in contrast to bilinear interpolation of spatial tables which smooths the learned structure; and (iii) it shares the same parameterization style as a frequency-domain spectral mixer, enabling co-design of positional encoding and token mixing in a common substrate. We demonstrate the bias inside our **Dual-Domain Residual (DDR)** framework, in which a spectral mixer and a gated attention head act as parallel residuals within each block. Head-to-head, the frequency-domain bias recovers $+1.98$ pp over Swin RelPos and $+2.74$ pp over no positional bias; removing the entire attention branch (which carries the bias) drops accuracy by $5.35$ pp, while removing the spectral branch drops it by $2.22$ pp. On ImageNet-1K from scratch, DDR outperforms SpectFormer at every comparable parameter budget, by $+2.0$ pp at Tiny (DDR-Ti-SF, 10.7M, 78.9\%) and $+0.4$ pp at Base (DDR-B-Deep, 63.1M, 82.5\%); at sub-parity parameters, DDR-Ti+ (8.9M, 77.9\%) exceeds SpectFormer-Ti (9.0M, 76.9\%) by $+1.0$ pp. The learned per-layer gates concentrate spectral processing in early blocks and attention in late blocks, qualitatively matching the schedules that are hand-coded in the previous hybrid models.

Large multimodal models (LMMs) have advanced rapidly in perception and reasoning, yet it remains unclear whether they can discover visually grounded, physically feasible solutions in open-ended environments. We first introduce MM-CreativityBench, a benchmark for affordance-grounded creative tool use, where models must inspect scenes, entities, and parts to identify non-obvious object uses grounded in visual evidence. Our evaluation shows that current LMMs often fail to sustain grounded exploration: they overlook relevant entities, under-examine critical parts, or hallucinate unsupported attributes. To address these failures, we propose affordance-grounded alignment, framing creative tool use as a preference learning problem. Using Direct Preference Optimization and supervision from an affordance knowledge base, we train models to favor visually grounded attribute–affordance reasoning over hallucinated alternatives while improving exploration efficiency. Our results yield consistent gains in correct entity and part selection, and reduce hallucination and grounding errors. These findings position grounded creativity as a core capability for future multimodal agents: the ability to adapt to unfamiliar environments and solve problems beyond memorized patterns or surface-level plausibility, moving closer to real human-like intelligence.

Value-model-free RLVR methods such as GRPO assign uniform advantages to all tokens in a rollout, ignoring that tokens contribute unequally. Recent methods use token entropy as an importance proxy but compute it globally across the batch, conflating importance with prompt difficulty and positional trends. We argue that importance should instead be measured relative to a token's own local context, and introduce proximal entropy: a local measure of token importance relative to neighboring tokens, and prove it is invariant to both confounders. Proximal Entropy Policy Optimization (PEPO) uses it to weight per-token advantages and outperforms GRPO and entropy-based baselines on mathematical reasoning across Qwen3-1.7B, Qwen3-4B, and Llama-3.2-3B-Instruct. We also show the formulation generalizes to other algorithms where substituting proximal entropy into existing methods improves, and applying it to single-stream RL succeeds where global entropy fails.

As large language models (LLMs) become increasingly deployed as autonomous agents, the ability of these models on automated planning tasks has become essential. However, despite their strengths, LLMs' planning ability is not robust, particularly when they are not allowed to exploit commonsense cues in the symbolic vocabulary (e.g., names of actions, predicates, objects) used to define planning tasks. This makes them unreliable compared to symbolic planners. With recent reasoning models showing improved planning performance on tasks with obfuscated symbolic vocabularies, we utilize an adversarial attack framework to reveal to what extent these crucial weaknesses of LLM agents remain and whether they can reason properly when presented with severe cases. As such, our main contribution is devising and optimizing an LLM-based attack model which we call $\texttt{Symbolic-Swapper}$ to automatically generate obfuscated domains with non-standard symbolic vocabularies that systematically break strong reasoning models' ability to plan. In doing so, we reveal by how much LLMs planning abilities can be further degraded under more advanced obfuscation schemes, allowing us to ascertain their success when having to rely on their pure planning ability. We find that attacks can decrease the planning performance of base models and even robust planning frameworks by over 70\%. We further analyze the factors behind how their planning abilities break down, and find that successful obfuscation models to significantly degrade in reasoning quality despite them showing an ability to understand the domain, revealing a gap in how model's perceive a task domain and how they are actually able to plan under it, with attacks often bypassing model ability to successfully self-correct.

Ensuring safety guarantees in offline reinforcement learning remains challenging, especially when safety constraints must hold almost surely. Moreover, as pre-specifying a single safety budget (constraint threshold) is often challenging, it is desirable to learn a foundation policy that can be deployed across a broad range of budgets. We introduce AEGIS (Almost-sure Epigraph-Guided Implicit Safety), a safe offline RL framework that can guide diffusion policy training via critics that respect almost-sure constraints across various feasible budgets. AEGIS characterizes the feasible set of initial state-budget pairs as the epigraph of a feasibility critic updated via the worst-case backup. Building on the proposed characterization, we extend implicit Q-learning (IQL) to train both feasibility and reward critics. We use these critics to bias a diffusion policy toward high-value feasible actions. Consequently, AEGIS turns diffusion from a generative prior into a safety-aware controller, enabling a single general policy to respect various budgets without further tuning. Empirical results on the DSRL benchmark and humanoid locomotion tasks show that AEGIS achieves high feasibility with competitive returns, generalizing across feasible constraint thresholds.

Modern gradient boosting algorithms are based on Newton's method, which excels empirically but lacks global convergence guarantees. While recent work on Gradient-Regularized Newton (GRN) boosting achieved a global $\mathcal{O}(1/\gamma^{12}k^2)$ rate, its analysis was not able to exploit the natural Hessian geometry and suffers from a $\gamma^{12}$ dependence on a worst-case `weak gradient edge' $\gamma$. In this paper, we tackle these theoretical drawbacks by generalizing the Affine-Invariant Cubic Newton (AICN) scheme to convex optimization with weak learners. Our analysis is based on average cosine values $\overline{\Theta}_k$ along the optimization path and directional improvements $\delta_k$ in the Hessian-induced norm. We show that the weak-AICN update reduces to a damped weak-Newton step with an explicit stepsize determined by $\delta_k$. Under semi-strong self-concordance, we establish a global $\mathcal{O}(1/\overline{\Theta}_k^4 k^2)$ rate, significantly improving upon the weak-GRN bound. We further prove a local linear rate, global linear convergence for Hessian-dominated losses, and cosine angle lower bounds. Finally, we demonstrate the broader applicability of this framework beyond gradient boosting by showing that Newton coordinate descent emerges naturally as a special case of optimization with weak learners.

Many everyday robot manipulation skills are affordance-dependent, with success determined by whether the robot contacts the functional object region required by the subsequent action. Current simulation data generators obtain contacts from generic grasp estimators or per-object manual contact annotations, but generic estimators rank stable grasps without task semantics and often select contacts that are misaligned with the downstream action, while manual contact annotations must be rewritten for each new object and task. To solve these challenges, we introduce AffordSim, a scalable data generator and benchmark that integrates open-vocabulary 3D affordance prediction into simulation-based trajectory generation. Given a natural-language task description, AffordSim synthesizes a task-relevant scene, emits affordance queries, grounds them on object surfaces, samples region-conditioned grasps, and selects executable candidates with motion planning. It further randomizes object pose, texture, lighting, image noise, and cross-viewpoint backgrounds for sim-to-real transfer. We instantiate AffordSim as a 50-task benchmark across diverse manipulation skills, five robot embodiments, and 500+ rigid and articulated objects. AffordSim achieves $93\\%$ of the trajectory collection success rate of manual contact annotations on affordance-critical tasks and $89\\%$ on hard composite tasks. Vision-language-action policies trained on AffordSim data transfer zero-shot to a real Franka FR3, reaching $24\\%$ average success. The data generator, benchmark, and code are available at \href{https://anonymous.4open.science/r/AffordSim-8D94/}{project page}.

We introduce Agent$^2$ RL-Bench, a compact diagnostic benchmark for evaluating agentic RL post-training, which tests whether LLM agents can autonomously design, implement, debug, and execute post-training pipelines that improve foundation models. RL post-training increasingly drives model alignment and specialization, yet existing benchmarks are largely static, rewarding supervised fine-tuning or script generation without assessing an agent's ability to close an interactive RL loop. Agent$^2$ RL-Bench provides a unified agent-facing interface: each run starts from an isolated workspace containing a base model, task data, instructions, and a grading API, and agents must iterate within a fixed budget by training models and submitting artifacts for evaluation. The benchmark spans six tasks across three levels, from static rule-based training to judge-based optimization and closed-loop online RL with trajectory collection. Two diagnostic skills, namely runtime recording and post-hoc summarization, enable structured analysis of agent behavior, facilitating smooth and effective iteration of the benchmark's evaluation framework. Across five agent systems and six driver LLMs, agents show intelligent behavior but clear limitations: one RL-oriented run improves ALFWorld from 4.85 to 93.28 via SFT warm-up and GRPO with online rollouts, yet DeepSearchQA remains difficult, most successful routes rely on supervised pipelines, and interactive outcomes show large single-run differences across agent stacks. Overall, Agent$^2$ RL-Bench shows that current agents can sometimes engineer online RL, but stable agent-driven RL post-training remains rare under fixed budgets. It also demonstrates that our benchmark provides a strong and effective evaluation framework for future research in this direction. Code is available at https://anonymous.4open.science/r/Agent2RL-Bench-5587.

Vision-language models with extended reasoning succeed on complex problems, but many real-world problems require external tools that internal reasoning alone cannot reach. Agentic reasoning therefore interleaves two behaviors with a structural asymmetry: thinking (the self-contained default) and tool use (a high-variance auxiliary acting). We refer to this asymmetry as the Thinking-Acting Gap. Under standard RL recipes like GRPO, the gap manifests as two diagnostic symptoms during training: tool use is attempted on only ~30% of rollouts, and when attempted, the tool-using rollouts within a group are all-wrong on ~40% of questions, suppressing the learning signal at the tool calls that needed it. We propose AXPO (Agent eXplorative Policy Optimization): for each all-wrong tool-using subgroup, AXPO fixes the thinking prefix and resamples the tool call and its continuation, paired with uncertainty-based prefix selection. Across nine multimodal benchmarks and three scales of Qwen3-VL-Thinking, SFT+AXPO outperforms SFT+GRPO at average (+1.8pp Pass@1 and +1.8pp Pass@4 at 8B on average) and 8B with SFT+AXPO surpasses the 32B Base on Pass@4 with 4x fewer parameters.

Acquiring high-quality labeled data and engineering effective learning pipelines are two critical bottlenecks to applying machine learning systems in new domains. Although recent MLE agents have shown promise in automating pipeline design, their behavior has been studied almost exclusively in fully supervised settings. In many domains, however, labeled data is expensive while unlabeled data is abundant. In this work, we study agents in the semi-supervised learning (SSL) regime by framing SSL as a program search problem and exploring how agents navigate this space under limited supervision. Our analysis focuses on three questions: (1) whether agents can effectively leverage unlabeled data, (2) how they leverage it and what strategies they discover, and (3) whether they can adapt these strategies to specific domains. We find that agents can leverage and synthesize SSL strategies spanning decades of SSL research, achieving performance comparable to, and in some cases outperforming, established human-engineered SSL methods. We further validate these observations on real-world ecological datasets, highlighting both the challenges and opportunities of transferring agent-discovered strategies beyond controlled benchmarks.

Temporal sequences in working memory are hierarchically organized to support flexible behavior. Recent neural recordings suggest that this organization relies on a 1D-to-2D neural geometrical folding in working memory, whereby ordinal positions are re-encoded within an orthogonal scaffold, with two dimensions indexing global (chunk-level rank) and local ranks (within-chunk position), respectively. This findings implicate an important structural principle for memory, yet the computational mechanisms underlying this geometry remain unclear. Here, we study this question using a recurrent neural network with fast Hebbian plasticity (Hebb-RNN). When meta-trained on rank recalling tasks, the Hebb-RNN generalizes to novel item–rank bindings and hierarchical sequences, outperforming control models. Mechanistic analysis reveals that the network develops an orthogonal ordinal geometry consistent with empirical findings. This geometry transforms temporal inputs into abstract ordinal ranks that provide a memory scaffold, while Hebbian plasticity supports rapid conjunctive binding of item content to this structure. Furthermore, the pre-learned geometric scaffold significantly accelerates learning in novel tasks, supporting efficient transfer. By incorporating a predictive module, our model extends to real-world encodings and acquires hierarchical geometry without explicit chunking signals. Together, our results provide a biologically plausible framework for hierarchical sequence representation, offering new insights into structured memory organization for both neuroscience and machine learning.

AI models deployed in critical domains, such as AI safety research, may subtly sabotage our efforts due to misalignment. Sandbagging is a form of sabotage in which an AI intentionally underperforms on the task given to it. It is particularly pernicious on fuzzy tasks, i.e. tasks which are hard to grade or require intuition. To understand sandbagging on fuzzy tasks, we introduce a novel AI control framework that considers sandbagging as an adversarial game between a blue team and a red team. The blue team uses a weak trusted model to construct a weak score against which they would train a strong, potentially sandbagging model to remove the sandbagging if it were present. The red team then tries to find model behaviors that are rated highly by the weak score, and thus might not be trained out, but actually correspond to poor performance. We test our framework on the task of writing experimental proposals for research questions from recent ML papers. We use a language model with access to the original paper as a proxy "ground-truth" scorer. Our red team discovers sandbagging behaviors using multi-objective evolutionary prompt optimization. We show that Opus 4.6 can write proposals that are worse according to the ground truth proxy than those of GPT-OSS-20B, while the weak scorer rates them as highly as the best proposals from Opus 4.6. To mitigate the threat of sandbagging, we propose an adversarial optimization algorithm for the blue team that discovers more robust prompts for the weak model. This algorithm produces a blue team prompt that our red team optimization fails to exploit.

Multimodal large language models offer a promising foundation for industrial anomaly detection (IAD), yet they remain unreliable for fine-grained diagnosis, where subtle defects require tight alignment between localized visual evidence and domain-specific semantics. Existing single-pass frameworks produce predictions without revisiting or validating intermediate evidence, which leads to spatio-semantic misalignment and error accumulation. More importantly, they lack mechanisms to actively control how additional evidence is acquired and integrated during reasoning, limiting their ability to resolve ambiguous cases. To address these limitations, we propose AirIAD, an agentic iterative reasoning framework that unifies progressive refinement with active information acquisition. AirIAD equips the model with two tools, Perceptive Zoomer and Knowledge Retriever, which enable it to iteratively gather complementary visual and semantic evidence. Through this process, the model performs explicit cross-verification between what is observed and what is known, refining intermediate hypotheses until convergence on a reliable diagnosis. This capability is enabled by a two-stage training pipeline. A supervised fine-tuning stage introduces a Spatio-Semantic Cross-Verification Chain-of-Thought, which structures coordinated perception and knowledge grounding. This is followed by a Spatio-Semantic Group-in-Group Policy Optimization stage, which provides dense step-wise supervision to encourage consistent self-refinement and robust spatio-semantic alignment. Extensive experiments on the MMAD benchmark show that AirIAD, built on Qwen3-VL-Instruct-4B, achieves state-of-the-art performance in anomaly detection and fine-grained diagnosis.

Contextual queueing bandits provide a framework for learning to schedule heterogeneous jobs under unknown context-dependent service rates. Under stochastic contexts, existing algorithms achieve $\widetilde{\cal O}(T^{-1/4})$ queue length regret, defined as the expected difference between the learner's and oracle's queue lengths at horizon $T$. In this paper, we improve this rate to $\widetilde{\cal O}(T^{-1/2})$. The key observation is that random exploration is needed only up to a carefully chosen cutoff round, rather than throughout the entire horizon. We propose CQB-$\eta$-2, a three-phase algorithm: (i) pure random exploration to construct an initial estimator, (ii) $\eta$-random exploration combined with a UCB rule to continue learning while maintaining negative drift, and (iii) pure UCB after the exploration cutoff. Our proof decomposes the queue length regret at the cutoff round. Before the cutoff, negative drift suppresses queue length differences caused by suboptimal choices. After the cutoff, the first two phases provide sufficient random exploration samples, ensuring that UCB decisions incur small departure-rate gaps. Combining these two bounds yields queue length regret of order $\widetilde{\cal O}(T^{-1/2})$. We further prove a minimax lower bound of order $\Omega(T^{-1/2})$. The proof constructs two hard instances that are statistically indistinguishable up to the final service decision, and uses a queue-specific coupling argument to convert the resulting testing error into queue length regret. Together, our upper and lower bounds characterize the minimax dependence on the horizon $T$ up to logarithmic factors.

The remarkable success of modern AI has been closely tied to scaling laws, yet the finite supply of high-quality data makes data efficiency—learning more from less—an increasingly important frontier. A model’s inductive bias is a critical lever for data efficiency, but foundational sequence models such as State Space Models (SSMs) often rely on fixed, task-agnostic biases. When this fixed prior is misaligned with the underlying structure of a task, the model may require additional samples to overcome its own bias before learning the relevant signal. In this work, we introduce a principled framework for understanding and aligning the inductive bias of linear time-invariant SSMs. We first formalize this bias through an SSM-induced kernel and show theoretically and empirically that its spectrum is governed by the model’s frequency response. This characterization motivates Task-Dependent Initialization (TDI), a fast power-spectrum matching method that aligns the initial SSM bias with the task’s spectral characteristics before downstream training. Across controlled synthetic experiments, trainable one-layer SSMs, and deep SSMs on diverse real-world benchmarks, TDI can improve data-efficient generalization primarily when task-relevant spectral structure is present and the default SSM bias is spectrally mismatched. Our results provide both a theoretical lens and a practical tool for task-adaptive inductive bias, suggesting a path toward more data-efficient sequence modeling.

Training large language models (LLMs) as multi-turn conversational agents remains a significant challenge, particularly in goal-oriented settings. The difficulty stems from sparse, long-horizon objectives and the discrepancy between response-level planning and token-level generation. In this paper, we present a formal reduction of the multi-turn RL problem into a \emph{sequence of single-turn RLHF-style problems}. This is achieved by setting a learned multi-turn Q-function as the reward model for the single-turn problem. We demonstrate and prove a key insight: solving this single-turn RLHF problem with standard token-level GRPO is equivalent to an approximate policy improvement step within the multi-turn problem. This insight naturally leads to \emph{Iterative GRPO}, a batch online approximate policy iteration algorithm that alternates between collecting a batch of data from the current policy, fitting Q-functions from these logged conversation trajectories, and improving the policy via single-turn RLHF. A major practical advantage is that Iterative GRPO directly leverages stable, off-the-shelf single-turn RLHF tools, making it straightforward to implement. In addition, the method occupies a middle ground between fully online and fully offline approaches, retaining the adaptability of online updates while gaining the stability benefits of offline training. Empirically, we demonstrate the effectiveness of Iterative GRPO on six multi-turn conversational environments.

Key--value (KV) cache quantization is widely used to reduce Large Language Model (LLM) inference memory, yet existing evaluations solely focus on measuring perplexity and accuracy without assessing the safety impact. In this study, we explore alignment preservation under KV cache quantization. Across eleven instruction-tuned models (3.8B--72B) and five benchmarks (1,894 prompts), we find that low-bit quantization can silently destroy safety alignment (Mistral-7B loses 15.2\% of its refusals at only $1.03\times$ perplexity, and no universal safe bit-width exists), with sharp model-specific phase transitions invisible to standard metrics. We identify that the root cause is geometric: safety features occupy a low-dimensional activation subspace $10^2$-$10^3\times$ more vulnerable to quantization noise than the full representation space perplexity averages over. Inspired by this observation, we propose **Per-Channel Reduction** (PCR), a diagnostic that classifies each model into one of three mechanistic failure modes: (i) *outlier-crushes-safety*, where safety lives in non-outlier channels collaterally damaged by outlier-driven scale factors; (ii) *outlier-as-safety*, where safety overlaps outlier channels and finer granularity cannot rescue it; (iii) *multi-layer dilution*, where safety is distributed across many layers and per-layer fixes fail. PCR predicts the correct mitigation direction on all nine primary models and one held-out model from an independent family (20 calibration prompts). PCR generalizes across unseen prompts, models, and production quantizers (KIVI: up to 97.2\% recovery), succeeding where attention-based allocation methods fail. The resulting training-free protocol (${\sim}35$ GPU-minutes) recovers up to 97\% of lost alignment at minimal memory overhead, addressing vulnerabilities confirmed in production vLLM serving with FP8 KV cache on NVIDIA GPUs.

Multimodal Foundation Models (MFMs) have made substantial progress, yet remain fragile in spatial reasoning over the physical world. A key bottleneck lies in their inability to transform local egocentric observations into a global allocentric spatial representation. To address this, we propose AlloSpatial, an agentic framework for allocentric spatial cognition in foundation models. AlloSpatial introduces World2Mind, a plug-and-play cognitive mapping sandbox that converts egocentric observations into structured allocentric priors, including Allocentric-Spatial Trees and route maps that support querying object topology, geometric relations, passability, and trajectories. To utilize these priors reliably under noisy reconstruction and ambiguous visual evidence, AlloSpatial introduces a Spatial Reasoning Harness for tool-use judgment, modality-decoupled cue collection, and geometry-semantic arbitration. We further internalize this process in Qwen3-VL through cold-start reinforcement learning with a harness-gated trajectory-level reward. Experiments on VSI-Bench and MindCube show that AlloSpatial improves proprietary models by 5\%-18\% in a training-free setting, while ASTs alone support strong spatial reasoning even when visual inputs are removed. The trained AlloSpatial agents further outperform larger general-purpose models and competitive spatial baselines, suggesting that structured allocentric representations, active tool use, and verifiable reasoning offer a promising route toward spatially capable foundation models.

Amorphous (disordered) materials require simulation cells with at least hundreds to thousands of atoms to accurately capture their diverse local atomic environments, compared to crystalline materials that can be described by unit cells containing a few to hundreds of atoms. To advance the design of amorphous materials with desired properties and facilitate the exploration of their vast design space, generative inverse design has emerged as a promising approach. It aims to directly output materials with properties closely aligned with the desired ones using probabilistic generative models conditioned on desired properties, which can be more resource efficient than the traditional trial-and-error approach. However, due to the inherent stochasticity of probabilistic generative models, when element assignments are unconstrained, a large portion of generated materials may be charge unbalanced, and no existing methods can effectively mitigate this limitation. In this work, we propose AMGenC, a new generative inverse design method for amorphous materials that can guarantee the generation of charge balanced samples, with minimal additional computational overhead and without sacrificing inverse design accuracy. AMGenC achieves this through an element noise that gives the generation process a starting point centered around charge balance, and the combination of a per-step soft projection and a final discrete projection for steering the elements toward exact charge balance throughout the generation. We perform extensive experiments on two amorphous materials datasets. Experimental results provide evidence that AMGenC achieves its design goal.

Language models often produce high-quality individual samples but poor sets, with repeated samples clustering around the same semantic modes. We formalize diverse generation as sampling a size-$K$ subset of complete strings from a determinantal point process (DPP), where the likelihood encodes item quality, and the embedding geometry encodes repulsion between similar outputs. Exact inference over all strings is intractable, so we propose \algname: draw a finite candidate pool from a tractable proposal, importance-weight candidates toward a globally tempered quality distribution, and run either $K$-DPP sampling or greedy MAP selection on the induced pool kernel. We prove a finite-sample bound showing that the KL bias of the pooled sampler decays as $\mathcal{O}(1/N)$ in the pool size. Empirically, \algname improves the quality--diversity Pareto frontier over temperature sampling, diverse beam search, $K$-means reranking, and LLM-based selection on open-ended generation, ambiguous text-to-SQL (Ambrosia), and mathematical reasoning (GSM8K) tasks.

Most modern deep neural networks are trained with Adam and related coordinate-wise adaptive optimizers, which are efficient but do not respect the natural matrix symmetries, geometry, and spectral structure of neural network parameters. We develop a symmetry-based principle for matrix-gradient optimization, showing that orthogonally equivariant optimizers, especially spectral optimizers such as stochastic spectral descent, Muon, Scion, and polar gradient methods, are better aligned with matrix layers than coordinate-wise updates. We show that coordinate-wise adaptivity breaks orthogonal equivariance, discards gradient spectral structure, and can be viewed as applying spectral updates to a pathological diagonal lifting of the parameter space. Motivated by permutation and orthogonal symmetries, we derive symmetry-compatible optimizers for different layer types, including linear layers, embedding and LM head matrices, and MoE routers. This yields one-sided spectral optimizers, row-norm optimizers, and hybrid row-norm/one-sided-spectral variants. Pre-training experiments on dense and sparse MoE language models, including Qwen3-0.6B-style, Gemma 3 1B-style, OLMoE-1B-7B-style, and downsized gpt-oss models, show that replacing AdamW on large vocabulary-indexed matrices and router matrices with symmetry-compatible optimizers consistently improves final validation loss and sometimes training stability.

Human motion change is currently hindered by the lack of large-scale datasets, comprehensive benchmarks, and robust algorithms capable of handling diverse motion requirements. To address these gaps, we propose AnyMotion, a novel open-world paradigm for highly controllable and identity-consistent human motion editing across arbitrary scenes and actions. Our contributions are two-fold. First, we establish a hierarchical label taxonomy covering six major categories and 527 fine-grained classes, and curate HME-1M, a large-scale dataset containing 1.04 million editing quadruplets via systematic data balancing, cleaning, and reverse construction. Second, we introduce the Skeleton Chain-of-Thought (SCoT) framework, a two-stage pipeline consisting of motion reasoning and guided generation. In the reasoning stage, a multimodal large language model serves as a cognitive planner to derive target skeletal keypoints and MetaQueries. These outputs subsequently function as explicit structural and semantic priors, providing the necessary guidance for the diffusion transformer to perform high-fidelity motion generation. Finally, built upon our label taxonomy, we present Human Motion-Bench, the first motion-centric benchmark equipped with multi-category fine-grained evaluation metrics. Extensive experiments demonstrate that AnyMotion achieves state-of-the-art performance across diverse editing scenarios.

Complex Instruction Following (CIF) is a foundational capability for deploying Large Language Models (LLMs) in production-grade applications. It demands stable, precise outputs under increasing horizon length, deeply nested rules, and multi-turn interactions. Yet existing benchmarks are predominantly atomic and short-horizon, failing to capture the logical depth and distractor pressure of real-world settings. We introduce **AppCIF**, the first application-level CIF benchmark. AppCIF stratifies difficulty along two orthogonal axes: *Horizon Length* ($H$), the span over which distractors accumulate and critical signals must be tracked; and *Logical Depth* ($L$), the nesting depth of rules to be unfolded. Together, $\langle H, L\rangle$ defines four hierarchical complexity levels, scaling from clean few-step execution (L1) to extreme pressure deduction (L4). AppCIF comprises 466 curated samples across five application scenarios. It adopts a "**Prompt-as-Environment**" paradigm, pairing Complex System Prompts (averaging ~14k words) with executable Code Simulation Engines (averaging ~3.5k lines of Python) to sustain rigorous state-machine deduction. The benchmark is constructed via a **Labor-Divided Human-AI pipeline** with scenario-specific collaboration strategies, and evaluated on 12 representative LLMs. Results reveal sharp performance stratification: Hard Satisfaction Rate ranges from 1.1% to 45.9%, exposing severe bottlenecks in long-horizon, distractor-laden instruction execution.

Operator learning has been highly successful for continuous mappings between infinite-dimensional spaces, such as PDE solution operators. However, many operators of interest—including differential operators—are discontinuous or set-valued, and lie outside classical approximation frameworks. We propose a paradigm shift by formulating approximation via graph convergence (Painlevé–Kuratowski convergence), which is well-suited for closed operators. We show that uniform and $L^p$ approximation are fundamentally inadequate in this setting. Focusing on maximally monotone operators, we prove that any such operator can be approximated in the sense of local graph convergence by continuous encoder–decoder architectures, and further construct structure-preserving approximations that retain maximal monotonicity via resolvent-based parameterizations.

Events in spatiotemporal systems are ubiquitous, yet modeling their complex distributions remains challenging. Existing point process models often rely on strong structural assumptions and are typically limited to autoregressive, event-by-event prediction. As a result, they struggle to support broader inference tasks such as inverse inference, trajectory reconstruction, and recovery of missing event locations. We introduce Arbitrarily Conditioned Hierarchical Flows (ARCH), a hierarchical flow matching framework for spatiotemporal event modeling. ARCH is expressive enough to capture complex event distributions while enabling tractable and accurate computation of conditional intensities, which quantify instantaneous event risk. Built on a history-encoder-generative-decoder architecture, ARCH introduces a hybrid masking strategy for flexible conditioning on arbitrary observed events. This enables a unified treatment of forecasting, inverse inference, and partial trajectory recovery within a single framework. Experiments on synthetic and real-world datasets show that ARCH consistently outperforms existing baselines across both prediction and conditional inference tasks.

We develop a reduction-based framework for online learning with delayed feedback that recovers and improves upon existing results for both first-order and bandit convex optimization. Our approach introduces a continuous-time model under which regret decomposes into a delay-independent learning term and a delay-induced drift term, yielding a delay-adaptive reduction that converts any algorithm for online linear optimization into one that handles round-dependent delays. For bandit convex optimization, we significantly improve existing regret bounds, with delay-dependent terms matching optimal first-order rates. For first-order feedback, we recover optimal regret bounds via a simpler, unified analysis. Quantitatively, for bandit convex optimization we obtain $O(\sqrt{d_{\text{tot}}} + T^{\frac{3}{4}}\sqrt{k})$ regret, improving the delay-dependent term from $O(\min\{\sqrt{T d_{\text{max}}},(Td_{\text{tot}})^{\frac{1}{3}}\})$ in previous work to $O(\sqrt{d_{\text{tot}}})$. Here, $k$, $T$, $d_{\text{max}}$, and $d_{\text{tot}}$ denote the dimension, time horizon, maximum delay, and total delay, respectively. Under strong convexity, we achieve $O(\min\{\sigma_{\text{max}} \ln T, \sqrt{d_{\text{tot}}}\} + (T^2\ln T)^{\frac{1}{3}} {k}^{\frac{2}{3}})$, improving the delay-dependent term from $O(d_{\textnormal{max}} \ln T)$ in previous work to $O(\min\{\sigma_{\text{max}} \ln T, \sqrt{d_{\text{tot}}}\})$, where $\sigma_{\text{max}}$ denotes the maximum number of outstanding observations and may be considerably smaller than $d_{\text{max}}$.

LLM-driven user simulators are increasingly adopted for A/B-testing recommender systems, yet the choice of backbone is treated as an interchangeable implementation detail rather than a factor shaping evaluation reliability. We show otherwise: across nine backbones under unified controls, this choice accounts for 149% as much outcome variance as the tested recommender, and swapping backbones alone flips about 20% of pairwise A/B winners. The backbone, in other words, is not a neutral component but the measurement instrument. We turn this observation into a method: ARES (Audit for Reliable Evaluation of Simulators) separates a backbone effect (the instrument's contribution) from a recommender effect (the signal under study), quantifies both at three levels (outcomes, trajectories, and reasoning), stress-tests them against text vs. rendered UI, and endorses only A/B claims a majority of backbones independently support. Applied to nine backbones $\times$ five recommenders on MovieLens-1M, ARES surfaces failures invisible to outcome metrics: within a single model lineage, newer and stronger releases do not necessarily agree more with their predecessors (Kendall's $\tau$ drops from 1.0 to 0.20 across generations); three of eight vision-capable backbones reverse the declared best recommender under UI rendering; and 63000 reasoning traces tie these reversals to interpretable behavioral biases, not stochastic noise. Capability alone, therefore, does not guarantee reliability. While instantiated for recommendation, the backbone-as-instrument view generalizes to any LLM-mediated evaluation where model identity is a free variable, including LLM-as-a-Judge. We release ARES-Bench at https://github.com/neurips2026-ares-authors/ARES-Bench: 21400 behavioral logs, 63000 reasoning traces, a reusable visual sandbox, and an analysis toolkit, making multi-backbone auditing a feasible default for simulator-based evaluation.

Watermark robustness under paraphrase is usually reported as AUC against a named rewriter, making results hard to compare across watermark schemes, paraphrasers, and detectors. We propose a retained-signal reporting interface: if a watermark injects KL signal $\Delta=\KL(P_1\Vert P_0)$ and a paraphrase channel retains KL fraction $\rho$, then any detector's post-paraphrase advantage is controlled by $\Delta\rho$. We give explicit Pinsker and Bretagnolle--Huber threshold laws, prove the $\Theta(\sqrt{\Delta\rho})$ rate is sharp for any $(\Delta,\rho)$-only converse, and show the bound is implementation-blind once watermark families are matched by $\Delta$. For real LLM paraphrases, we instantiate the interface as a score-projected audit that estimates $\widehat{\Delta}\widehat{\rho}$ for a deployed detector and reports it with semantic preservation and the detector projection. In a KGW/Unigram $\times$ Qwen/Llama benchmark, this score-effective signal predicts post-paraphrase AUC across lengths, strengths, and paraphraser families. The resulting protocol separates injected signal, retained signal, semantic quality, and detector choice, replacing one-off AUC tables with a reproducible robustness report.

Identity-preserving video generation is fundamentally constrained by the prevailing point-reference paradigm: existing methods consume one or a few snapshot images and therefore cannot transcend "look-alike" fidelity, suffer from severe drift under large yaw, occlusion or scale variation, and trade brittle cross-pair supervision against pervasive reference leakage. We argue that identity is intrinsically a spatially-distributed, multi-view concept and propose Argus, a theoretically-grounded framework that lifts identity injection from single-snapshot conditioning to multi-view aggregated reasoning. Argus comprises three tightly-coupled modules. (i) An Identity Director built on a frozen multimodal LLM automatically curates nine semantically-diverse keyframes, extracts a global identity descriptor, and rewrites the prompt to dissolve the prompt-frame-identity tri-conflict. (ii) A Stacked Mosaic Identity Injection packs the nine SMPL-localized crops into a 3×3 mosaic and feeds it through Wan 2.1's native 36-channel patch embedding via a flow-matching-synchronized prefix-token stream, while a Negative-Time-Anchored RoPE preserves architectural compatibility with Wan's split-axis design. (iii) An Adaptive Skip-Layer Guidance replaces hand-tuned PAG with a tiny mask network that learns where to perturb, decoupling identity guidance from textual guidance, and a Temporal Identity Annealing schedule injects identity only during the structure-formation window to obviate copy-paste artifacts. Trained without any cross-pair data, Argus attains state-of-the-art performance on OpenS2V-Eval, surpassing both closed-source (Hailuo) and open-source (VACE-14B, Phantom-14B, Stand-In, BindWeave) systems on five of seven metrics. To stress-test extreme failure modes, we additionally release HardID-Celeb, a celebrity benchmark with two new robustness probes (YawScore, OccScore) on which Argus establishes the highest scores by a wide margin. A 47-participant double-blind user study confirms perceptual superiority across both expert and lay populations.

Continual reinforcement learning challenges agents to acquire new skills while retaining previously learned ones with the goal of improving performance in both past and future tasks. Most existing approaches rely on model-free methods with replay buffers to mitigate catastrophic forgetting; however, these solutions often face significant scalability challenges due to large memory demands. Drawing inspiration from neuroscience, where the brain replays experiences to a predictive World Model rather than directly to the policy, we presentARROW (Augmented Replay for RObust World models), a model-based continual RL algorithm that extends DreamerV3 with a memory-efficient, distribution-matching replay buffer. Unlike standard fixed-size FIFO buffers, ARROW maintains two complementary buffers: a short-term buffer for recent experiences and a long-term buffer that preserves task diversity through intelligent sampling. We evaluate ARROW on two challenging continual RL settings: Tasks without shared structure (Atari), and tasks with shared structure, where knowledge transfer is possible (Procgen CoinRun variants). Compared to model-free and model-based baselines with replay buffers of the same-size, ARROW demonstrates substantially less forgetting on tasks without shared structure, while maintaining comparable forward transfer. Our findings highlight the potential of model-based RL and bio-inspired approaches for continual reinforcement learning, warranting further research. Code is available at https://github.com/Cerenaut/ARROW.

We consider time discretization for score-based diffusion models to generate samples from a learned reverse-time dynamic on a finite grid. Uniform and hand-crafted grids can be suboptimal given a budget on the number of time steps. We introduce Adaptive Reparameterized Time (ART), which controls the clock speed of a reparameterized time variable to redistribute computation along the sampling trajectory while preserving the terminal time, with the objective of minimizing the aggregate Euler discretization error. We derive a randomized companion ART-RL that recasts ART as a continuous-time reinforcement learning problem with Gaussian policies, and prove a two-directional bridge between the two: the deterministic ART optimum lifts to an optimal Gaussian policy, and conversely any optimal Gaussian policy must recover the ART control through its mean. This bridge turns continuous-time actor--critic learning into a principled, rather than heuristic, route to the deterministic timestep optimum. Within the official EDM pipeline, ART-RL improves FID on CIFAR--10 across a wide range of budgets; after one-time offline training, the distilled deterministic schedule transfers without retraining to AFHQv2, FFHQ, and ImageNet at no extra inference cost.

Aphasias, selective language impairments which can arise from brain damage, reveal the functional organization of human language by providing causal links between affected brain regions and specific symptom profiles. Drawing on this literature, we introduce an aphasia-inspired technique to characterize the emergent functional organization of language models (LMs). We ``lesion'' (zero-out) model parameters and measure the effects of this intervention against clinical aphasia symptoms, as diagnosed by the Text Aphasia Battery (TAB). When applied to 112,426 outputs from five 1B-scale LMs, the full range of evaluated symptoms surface, but in distributions largely distinct from those of humans. Our method uncovers broad, component-level symptom differences between attention (query, key, value, output) and feed-forward (up, gate, down) projections, without any significant within-class differences. We also find an effect of depth, where lesions in early layers disproportionately cause syntactic and semantic symptoms while late-middle layers yield higher rates of phonological and fluency deficits. Although some LM lesions induce quantitatively more similar profiles to some human aphasia types than others, qualitative differences in symptom patterns between LMs and humans suggest that aphasia syndromes are heavily influenced by the details of learning and processing rather than being a domain-invariant consequence of disrupted language processing.

Federated Learning (FL) is vulnerable to model poisoning attacks, where malicious clients manipulate uploaded model updates to corrupt the global training process. Existing attacks are typically evaluated by their eventual damage, such as final accuracy degradation or convergence to random-guess performance, where the learned model becomes unusable and may ultimately lead to denial-of-service (DoS). However, such metrics overlook attack round complexity, defined as the number of communication rounds required for the poisoned global model to reach and remain near a desired degradation objective. In realistic FL systems, requiring more communication rounds increases the adversary's exposure to client participation, robust aggregation, filtering, and detection. We propose ASAP (Adaptive Sliding Agnostic Poisoning), a fast aggregation-rule-agnostic (AGR-agnostic) attack that treats the whole FL training process as an uncertain dynamical system and formulates poisoning as a feedback-control problem for steering the model toward a desired attack objective. ASAP combines adaptive sliding mode control (ASMC) with finite Fourier-basis uncertainty estimation, treating the unknown effects of benign training and aggregation as a time-varying uncertainty and using the resulting estimate to design the malicious updates. We theoretically prove that the sliding variable reaches the sliding surface in finite time and the tracking error subsequently converges exponentially. Experiments across multiple datasets, models, and aggregation rules demonstrate that ASAP reaches specified degradation objectives in fewer communication rounds while maintaining smaller target deviation than existing model poisoning attacks.

Long-context inference with large language models is bottlenecked by the linear growth of the KV cache. Existing eviction-based methods reduce memory overhead by irreversibly discarding historical states, while retrieval-based methods preserve the full context but rely on retrieval units defined by tokens, fixed pages, or proxy signals including query-key similarity, clustering, and linguistic rules. Such externally defined units often misalign with the model's actual contextual dependencies during generation. We propose AsdaKV, an adaptive semantic drift-aware KV cache retrieval framework that defines semantic continuity directly from the model's inherent attention behavior. AsdaKV detects semantic drift by measuring the overlap between high-attention historical regions across decoding steps: stable overlap indicates that the active cache still covers the relevant contextual evidence, whereas a sharp overlap drop triggers cache refresh. Guided by this drift signal, AsdaKV organizes historical KV states into variable-length semantic windows, offloads full window KV states to CPU memory, and maintains only lightweight window anchors plus a fixed-budget active KV cache on GPU. During decoding, the current query retrieves relevant windows via anchor matching, while adaptive-stride drift detection and deferred page recall reduce redundant detection, retrieval, and KV reloading overhead. Extensive experiments across diverse tasks and model families demonstrate that under the same KV budget, AsdaKV achieves higher accuracy than state-of-the-art KV retrieval methods, maintains comparable decoding efficiency in long-input settings, and further outperforms existing retrieval systems for long-output generation.

Long-horizon embodied tasks remain a fundamental challenge in AI, as current methods rely on hand-engineered rewards or action-labeled demonstrations, neither of which scales. We introduce ASH, an agentic system that learns an embodied policy from unlabeled, noisy internet video, without reward shaping or expert annotation. ASH follows a self-improvement loop; when it gets stuck, ASH learns an Inverse Dynamics Model (IDM) from its own trajectories, and uses its IDM to extract supervision from relevant internet video. ASH uses unsupervised learning to identify key moments from large-scale internet video and retains them as long-term memory --- allowing it to tackle long-horizon problems. We evaluate ASH on two complementary environments demanding multi-hour planning: Pokemon Emerald, a turn-based RPG, and The Legend of Zelda: The Minish Cap, a real-time action-adventure game. In both games, behavioral cloning, retrieval-augmented and zero-shot foundation-model baselines plateau, while ASH sustains progression across our 8-hour evaluation. ASH reaches an average of $11.2/12$ milestones in Pokemon Emerald and $9.9/12$ in Legend of Zelda, while the strongest baseline gets stuck in both environments at an average of $6.5/12$ and $6.0/12$ milestones, respectively. We demonstrate that self-improving agents are a scalable recipe for long-horizon embodied learning.

Today, we improve models by training and evaluating them on problems at the frontier of their abilities. Creating such problems is itself a demanding task, requiring the ability to probe model limits and generalize beyond existing question distributions. It also means placing problems at a precise difficulty level, which requires understanding what it takes to solve them. In short, generating problems calibrated to a model's current frontier demands capability beyond it, an increasingly burdensome constraint as models improve. Our key insight is that we can leverage this constraint to our advantage: a model that can generate problems consistently calibrated to a given frontier must possess capability beyond it. Accordingly, we present Ask-E, an environment that benchmarks and trains models on their ability to write questions at a given skill level, rather than answer them. Concretely, we define target skill levels as ranges bounded by the capabilities of two existing language models. A generated question is successfully calibrated if exactly one of the two models can solve it, placing it precisely within the target range and differentiating the capabilities of these models. Ask-E serves both as a benchmark and a training environment, where models generate problems calibrated to a variety of skill levels. We find that even frontier models achieve below 50\% calibration on the benchmark, leaving significant headroom to measure future progress. We also show that training on this environment leads to improvements across a number of downstream math benchmarks even with no new math data, no interaction with stronger models, and no correctness-based reward.

Long-horizon AI agents execute complex workflows spanning hundreds of sequential actions, yet a single wrong assumption early on can cascade into irreversible errors. When instructions are incomplete, the agent must decide not only whether to ask for clarification but when, and no prior work measures how clarification value changes over the course of execution. We introduce a forced-injection framework that provides ground-truth clarifications at controlled points in the agent's trajectory across four information dimensions (goal, input, constraint, context), three agent benchmarks, and four frontier models (three per benchmark; one on a single benchmark only; 84 task variants; 6,000+ runs). Counter to the common intuition that ``earlier is always better,'' we find that the value of clarification depends sharply on what information is missing: goal clarification loses nearly all value after 10% of execution (pass@3 drops from 0.78 to baseline), while input clarification retains value through roughly 50%. Deferring any clarification type past mid-trajectory degrades performance below never asking at all. Cross-model Kendall τ correlations (0.78-0.87 among models sharing identical task coverage; 0.34-0.67 across the full 4-model panel) confirm these timing profiles are substantially task-intrinsic. A complementary study of 300 unscripted sessions reveals that no current frontier model asks within the empirically optimal window, with strategies ranging from over-asking (52% of sessions) to never asking at all. These empirical demand curves provide the quantitative foundation that existing theoretical frameworks require but have lacked, and establish concrete design targets for timing-aware clarification policies. Code and data will be publicly released.

Large language models are increasingly deployed as LLM-as-a-Service (LLMaaS), where user queries are sent to remote LLMs for inference, raising severe privacy concerns about the exposure of sensitive user inputs. Differentially private inference mitigates these risks by injecting noise into input embeddings. However, existing methods struggle to balance privacy and utility, as stronger privacy protection often leads to substantial accuracy degradation in LLM inference. In this work, we propose DP-SPD, a differentially private LLM inference framework that protects user queries by hiding the real query among client-generated dummy queries. All queries are randomly shuffled before being sent to the server, preventing direct identification of the user's true intent. To improve inference efficiency, we exploit the shared-prefix structure among queries and reuse transformer key--value (KV) caches during decoding. We analyze the privacy guarantee of DP-SPD and evaluate its utility on classification and generation tasks, as well as its privacy protection under EIA, MIA, and GPT-based attacks. Experimental results show that our method achieves strong inference-time privacy protection while maintaining utility close to plain-text inference under practical latency.

AdamW is ubiquitous in deep learning, yet its behavior remains poorly understood. We analyze its dynamics through the lens of dynamical systems and show that AdamW admits an \emph{implicit fixed-point objective}: its fixed points coincide with the stationary points of a constrained and regularized optimization problem. However, not all of these fixed points are stable under AdamW’s dynamics, and stability depends sensitively on curvature, weight decay, and momentum parameters. Even in simple one-dimensional settings, AdamW can exhibit surprisingly complex behavior: equilibria may be unstable, and numerical trajectories can exhibit persistent oscillations rather than convergence to them. We further extend the analysis to higher dimensions, deriving sufficient conditions for local stability of the continuous-time dynamics, and propose a convergent modification that preserves AdamW's fixed points. These results clarify what optimization problem AdamW is associated with, when its convergence can be expected, and how its dynamics could inspire more reliable optimizers.

As learning systems increasingly shape everyday decisions, Algorithmic Collective Action (ACA), i.e., users coordinating changes to shared data to steer model behavior, offers a complement to regulator-side policy and corporate model design. Real-world collective actions have traditionally been decentralized and fragmented into multiple collectives, despite sharing overarching objectives, with each collective differing in size, strategy, and actionable goals. However, most of the ACA literature focuses on single collective settings. To address this, we propose the first comprehensive statistical framework for ACA with multiple collectives acting on the same system. In particular, we focus on collective action in classification, studying how multiple collectives can influence a classifier's behavior. We provide quantitative statistical bounds on the success of the collectives, considering the role and the interplay of the collectives' sizes and the alignment of their goals. We make such bounds computable by each collective with only partial knowledge of other collectives' sizes and strategies. Finally, we numerically illustrate our framework on simulations inspired by interventions for climate adaptation in smart cities, demonstrating the usefulness of our bounds.

Large language model (LLM)-based agents have emerged as powerful autonomous controllers for digital environments, spanning mobile interfaces, operating systems, and web browsers. Web navigation, for example, demands handling dynamic content and long action sequences, making it a particularly complex task. Existing LLM-backed agents exhibit weakened long-horizon planning abilities during RL fine-tuning, where sparse and delayed rewards from terminal Outcome Reward Models (ORMs) make it difficult for agents to identify the actions that lead to success, preventing them from sustaining coherent reasoning over extended tasks. We address this with two contributions: (1) a validated subgoal generation procedure that produces reliable, monotonically calibrated progress signals between the initial state and the terminal ORM reward; and (2) MiRA ($\underline{Mi}$lestoning your $\underline{R}$einforcement Learning Enhanced $\underline{A}$gent), an RL training framework using dense, milestone-based reward signals via a learned potential critic. Starting from the open Gemma3-12B base ($6.4$%), MiRA enables the agent to reach $43.0$% on the WebArena-Lite benchmark. This performance surpasses proprietary systems such as GPT-4-Turbo ($17.6$%) and GPT-4o ($13.9$%), as well as the previous open-model state of the art, WebRL ($38.8$%). Our findings demonstrate that milestone-based reward shaping significantly boosts an agent's long-horizon abilities, paving the way for more robust, general-purpose autonomous systems.

Deep Q-Networks (DQNs) approximate value iteration by freezing a target network, forming Bellman-optimality targets, and running an inner regression loop that minimizes a smooth \emph{surrogate loss} (e.g., MSE/Huber loss) on data from a replay buffer. Yet, control is governed by the Bellman optimality operator's $\gamma$-contraction in $\ell_\infty$, whereas common objectives (MSE/Huber loss) measure progress in an averaged $\ell_2/\ell_1$ geometry; thus, improving the surrogate need not reduce the worst-case Bellman error that drives performance. To bridge this gap we connect the progress in the inner loop for each surrogate to the Bellman residual under the $\ell_\infty$ norm. Our analysis provides explicit thresholds denoted by $\alpha$ under which sufficient progress guarantees a $\ell_\infty$ contraction up to a standard approximation floor. This yields new convergence guarantees for DQN with fixed targets trained using MSE or Huber-type losses, and motivates a novel soft-$\ell_\infty$ surrogate that smoothly approximates the sup-norm and better matches the $\ell_\infty$ geometry resulting in the least stringent inner loop accuracy requirements. Across Atari benchmarks, soft-$\ell_\infty$ consistently reduces the worst-case Bellman residual and matches or improves returns relative to standard objectives, providing a practical, geometry-aligned recipe for stable DQN training.

Instruction-driven mask-free image editing with large-scale Diffusion Transformers (DiTs) has enabled high-fidelity local edits while maintaining global consistency. However, these models incur a prohibitive inference cost due to a pronounced semantic-–computation mismatch, where dense computation is performed over all high-resolution tokens regardless of localized editing instructions. Our analysis reveals that non-edit regions converge early with highly stable block-level outputs, exhibiting a distinct spatio-temporal asymmetry. Furthermore, we identify a critical system-level challenge that introducing such asymmetric computation into pipeline-parallel environments for ultra-large models triggers severe workload imbalance and pipeline bubbles, preventing algorithmic savings from translating into actual latency gains. To address these issues, we propose AsymPipe, a training--free algorithm–-system co-design framework that accelerates large-scale DiTs editing. The framework identifies edit versus non-edit regions on the fly using a zero-overhead velocity-based probe that reuses the predicted velocity field from the existing denoising process. For non-edit regions, AsymPipe caches and reuses block-level outputs, while for edit regions, it introduces a source-aware sparse attention module to precisely eliminate redundant contributions from static text and image sources. Finally, to bridge the gap between algorithmic optimization and hardware speedup, the framework incorporates a load-balanced token rearrangement scheme. This scheme deconstructs the original spatial topology to distribute tokens evenly across devices, physically restoring compute balance in pipeline-parallel systems. On HunyuanImage-3.0-Instruct and Qwen-Image-Edit-2511, AsymPipe achieves end-to-end speedups of 2.862x (PSNR 33.69) and 2.707x (PSNR 29.74), respectively, delivering substantial acceleration with virtually no loss in editing quality. Our anonymized source code is available at https://anonymous.4open.science/r/AsymPipe-6553.

Confidence sequences based on test martingales provide time-uniform uncertainty quantification for the mean of bounded IID observations without parametric distributional assumptions. Their practical efficiency, however, depends strongly on the choice of martingale updates, and many existing constructions do not exploit prior information about plausible data-generating distributions or mean values. We propose a Bayes-assisted framework that uses a Bayesian working predictive model to adaptively construct confidence sequences. For each candidate mean and time point, the predictive distribution selects, among valid one-step martingale factors, the update maximising predictive expected log-growth; validity is therefore preserved even when the prior or working model is misspecified. We prove that if the predictive distribution is Wasserstein-consistent, the resulting procedure is asymptotically log-optimal, matching the per-sample log-growth of an oracle procedure with access to the true distribution. We instantiate the framework using robust predictives based on Dirichlet-process mixtures and Bayesian exponentially tilted empirical likelihood. Experiments on synthetic data, sequential best-arm identification for LLM evaluation, and prediction-powered inference show that informative priors can substantially reduce confidence-sequence width and sampling effort while retaining anytime-valid coverage.

On-policy distillation (OPD) is an increasingly important way to post-train large language models (LLMs), but, like reinforcement learning (RL), it relies on rollouts from the model being optimized. For reasoning workloads, rollout generation can dominate training time. Asynchronous RL alleviates this bottleneck by decoupling rollout generation from learner updates, but doing so introduces stale-policy data; prior work studies how to stabilize learning from such data. However, it remains underexplored whether these asynchronous RL ideas and stale-data solutions transfer to OPD, and what OPD-specific constraints arise. To address this gap, we provide the first systematic study of staleness in asynchronous OPD. We first show that KL direction changes the stale-data problem: teacher-weighted forward KL is robust to stale rollouts, whereas student-weighted reverse KL is vulnerable. Second, for this vulnerable reverse-KL case, we study whether methods designed to stabilize asynchronous RL can mitigate OPD staleness. We find that they do not improve over a simpler OPD-specific surrogate: recomputing the reverse-KL signal under the current student at learner time without clipping. Third, we identify an OPD-specific teacher-cache constraint: under asynchronous execution, teacher scores are available at learner time only on cached actions. The resulting bias-variance tradeoff for sparse and sampled reverse-KL OPD implementations motivates multi-sample Monte Carlo (MC), which preserves MC correctability while reducing one-sample variance. Finally, we present and open-source **AsyncOPD**, a fully asynchronous OPD training pipeline built from these estimator choices. Experiments show that AsyncOPD improves training throughput by $1.6\times$ to $3.8\times$ over strict synchronous training while reaching comparable accuracy.

Small-molecule foundation models are typically pretrained on standalone molecular data, unlike vision and language models that often benefit from cross-modal or relational supervision. Protein-ligand co-folding provides a molecular analogue of such supervision by exposing models to atom-level ligand-protein interactions, raising the question of whether co-folding models can yield strong small-molecule representations. We study this question using Boltz2, a modern co-folding model, by transferring its atom-level ligand representations to standalone small-molecule tasks. Through systematic probing and distillation, we show that Boltz2 representations match or outperform existing models on the ADMET benchmark, accelerate molecular generative modeling, and improve sample efficiency in structure-guided ligand optimization. We further find that Boltz2 representations are complementary to those learned from conventional standalone molecular supervision, including 3D conformers, bioassay labels, and quantum-chemical properties. Finally, we extend representation alignment to reinforcement learning, showing that dense representation-level supervision can complement scalar rewards in molecular discovery. These results identify protein-ligand co-folding as a promising pretraining paradigm for small-molecule representation learning and position Boltz2 as a strong, off-the-shelf molecular foundation model.

Open-set 3D macromolecule detection in cryogenic electron tomography eliminates the need for target-specific model retraining. However, strict VRAM constraints prohibit processing an entire 3D tomogram, forcing current methods to rely on slow sliding-window inference over extracted subvolumes. To overcome this, we propose FullTilt, an end-to-end framework that redefines 3D detection by operating directly on aligned 2D tilt-series. Because a tilt-series contains significantly fewer images than slices in a reconstructed tomogram, FullTilt eliminates redundant volumetric computation, accelerating inference by orders of magnitude. To process the entire tilt-series simultaneously, we introduce a tilt-series encoder to efficiently fuse cross-view information. We further propose a multiclass visual prompt encoder for flexible prompting, a tilt-aware query initializer to effectively anchor 3D queries, and an auxiliary geometric primitives module to enhance the model's understanding of multi-view geometry while improving robustness to adverse imaging artifacts. Extensive evaluations on three real-world datasets demonstrate that FullTilt achieves state-of-the-art zero-shot performance while drastically reducing runtime and VRAM requirements, paving the way for rapid, large-scale visual proteomics analysis. All code and data will be publicly available upon publication.

Visual reasoning, often interleaved with intermediate visual states, has emerged as a promising direction in the field. A straightforward approach is to directly generate images via unified models during reasoning, but this is computationally expensive and architecturally non-trivial. Recent alternatives include agentic reasoning through code or tool calls, and latent reasoning with learnable hidden embeddings. However, agentic methods incur context-switching latency from external execution, while latent methods lack task generalization and are difficult to train with autoregressive parallelization. To combine their strengths while mitigating their limitations, we propose ATLAS, a framework in which a single discrete "word", termed as a functional token, serves both as an agentic operation and a latent visual reasoning unit. Each functional token is associated with an internalized visual operation, yet remains a standard token in the tokenizer vocabulary and can be generated via next-token prediction. This design avoids verbose intermediate visual content generation, while preserving compatibility with standard SFT and RL training, without architectural or methodological modifications. To further address the sparsity of functional tokens during RL, we introduce Latent-Anchored GRPO (LA-GRPO), which stabilizes the training by anchoring functional tokens with a statically weighted auxiliary objective, providing stronger gradient updates. Extensive experiments and analyses demonstrate that ATLAS achieves superior performance on challenging benchmarks while maintaining clear interpretability. We hope ATLAS offers a new paradigm inspiring future visual reasoning research.

We study the numerical approximation of Schrödinger Bridges in the setting where samples from the target distribution are unavailable and only an unnormalized target density is given. In this regime, the widely used iterative proportional fitting (IPF) procedure (and its variants) often exhibits severe numerical instabilities, stemming from ill-conditioned underlying control problems and from the fact that the optimal solution may lie far from the initialization. To address these issues, we propose a trust region variant of IPF that constrains successive updates to remain at a fixed Kullback-Leibler divergence from the current iterate. This restriction significantly improves numerical stability. Moreover, we show that the resulting method admits an interpretation as a mirror descent scheme with an adaptively chosen step size. We demonstrate the effectiveness of our approach on several challenging sampling tasks.

Audio-visual representation learning commonly aligns paired audio and visual signals with sample-level contrastive objectives. However, real-world audio-visual pairs usually contain heterogeneous patch tokens: only a subset carries reliable cross-modal shared semantics, while many others mainly encode modality-private residuals. Aggregating all tokens into a single global representation can therefore introduce semantic noise into cross-modal alignment. Motivated by the minimal-sufficiency principle, we propose an attribution-guided shared-private decoupling framework to mitigate this issue. Our key idea is to identify shared-cover subviews that preserve dominant cross-modal information while suppressing modality-private residuals, and use them to supervise decoupled shared and private representations. To identify such subviews efficiently, we draw on transformer attribution method and introduce Layer-Truncated Attribution (LTA), which provides a lightweight estimate of token-level relevance to cross-modal shared information. We further adopt a teacher-student framework, where the teacher provides attribution-induced supervision and the student internalizes shared-private decoupling into its representations, without introducing extra inference cost. Experiments on retrieval, classification, and sound-prompted semantic segmentation show consistent gains over strong baselines. Compared with baseline, our method improves the average zero-shot R@1 by 4.0 across AudioSet and VGGSound retrieval, and improves AS20K classification mAP by 3.1, demonstrating the effectiveness of fine grained attribution guidance for reducing semantic noise.

We introduce the metagame, a conceptual framework for quantifying second-order interaction effects of model explanations. For any first-order attribution $\phi(f)$ explaining a model $f$, we measure the directional influence of feature $j$ on the attribution of feature $i$, denoted as meta-attribution $\varphi_{j \to i}(f)$, by treating the attribution method itself as a cooperative game and computing its Shapley value. Theoretically, we prove that attributions hierarchically decompose into meta-attributions, and establish these as directional extensions of existing interaction indices. Empirically, we demonstrate that the metagame delivers insights across diverse interpretability applications: (i) quantifying token interactions in instruction-tuned language models, (ii) explaining cross-modal similarity in vision-language encoders, and (iii) interpreting text-to-image concepts in multimodal diffusion transformers.

Unifying speech, sound, and music generation in one model is hindered by tradeoffs between fidelity, end-to-end training, in-context conditioning, and variable-length synthesis that no current paradigm fully resolves. To address this challenge, we present AudioCALM, a universal audio generation framework that extends autoregressive (AR) next-token prediction from discrete tokens to continuous audio latents: a thin flow-matching head replaces the softmax to predict rectified-flow velocities at each position, and a block-causal AR-Flow attention pattern produces arbitrary-length output. Joint training of multiple audio generation tasks faces an asymmetric text--audio mismatch: speech transcripts align to specific time spans and demand \textit{tight, time-aligned} attention, whereas sound and music captions describe only overall semantics and rely on diffuse, holistic attention; mixing the two disproportionately degrades sound and music generation. We address this asymmetry at two levels: a \textbf{data reformulation} strategy that unifies all three tasks under a single description-style conditioning interface, and a novel architecture Asymmetric Mixture-of-Modality-Experts (A-MoME), which adds a dedicated residual expert for speech while sound and music share the backbone, incurring no inference overhead on non-speech inputs. Experimental results demonstrate that AudioCALM matches modality-specific state-of-the-art and outperforms prior unified baselines on speech, sound, and music generation benchmarks.

We present Audio-Visual Flamingo (AV-Flamingo), a fully open state-of-the-art audio-visual large language model (AV-LLM) for joint understanding and reasoning over audio, images, and long-form videos. Unlike prior AV-LLMs that primarily focus on short clips, AV-Flamingo is designed for understanding and reasoning over long and complex real-world (audio-visual) videos. To support this, we make three key contributions: (i) Audio-Visual-Skills, a large-scale collection of real-world videos with 7M caption and question-answer training instances designed to emphasize temporal, compositional, and cross-modal audio-visual reasoning; (ii) a novel three-stage curriculum that progressively trains the model from short-range perception to long-horizon multi-event reasoning; and (iii) Temporal Audio-Visual Interleaved Chain-of-Thought, a reasoning framework that explicitly grounds intermediate reasoning steps to timestamps in long audio-visual streams, improving temporal alignment and interpretability. Extensive experiments across 15+ audio-visual, omni-modal, audio, and vision benchmarks show that AV-Flamingo outperforms similarly sized open models by clear margins and remains highly competitive with, and in some cases surpasses, much larger open-weight and closed models, particularly on long and complex real-world audio-visual understanding and reasoning tasks. Beyond benchmark performance, AV-Flamingo exhibits strong real-world utility and transfers well to unseen tasks, highlighting its robustness and generalization ability. We will release the data, code, methods, and checkpoints upon acceptance. Project: https://avflamingo.github.io/

Public validation sets in machine-learning challenges frequently become de facto test sets in downstream work, making train–validation source overlap a first-order concern for claim validity. We audit the released public split of the Capsule Vision 2024 Challenge (CV2024) using perceptual hashing (pHash+dHash, PDQ at 99.4%), pixel NCC, and DINOv2/ResNet feature checks. The released CV2024 public split contains 1,381 / 11,581 KVASIR validation rows (11.9%; 540 pixel-confirmed at NCC ≥ 0.99, i.e. 4.66%) with pHash-exact training matches — directly violating the organizers' "NOT have duplicates" preprocessing description. All 1,381 trace to shared Kvasir-Capsule video prefixes, consistent with frame-level (not video-level) random splitting. The full CV2024-KVASIR slice falls in the disclosed Kvasir-Capsule near-duplicate set (60.7% bit-identical; non-KVASIR controls ≤ 0.31%). Training on the Kvasir-origin-removed le6 split drops public-validation balanced accuracy by Δ = −0.213 (95% CI [−0.220, −0.206], n=10 paired seeds), quantifying the public score's dependence on the KVASIR channel; the magnitude is mechanically expected given the 71.8% KVASIR validation composition rather than a leakage-specific causal effect. le6 is not a hidden-test proxy (team-level Spearman against AIIMS: ρ=0.355 vs. 0.566 for the original score). We release le6, le6plusinternal, hash/NCC annotations, audit code, and a Croissant 1.1 descriptor as a within-public-pool sensitivity probe; no image bytes are redistributed. Detector portability: ISIC 2019 cross-source rate is 0.008% (2/25,331, both re-attributed as within-source metadata-gap pairs ⇒ genuine cross-source 0/25,331; Cassidy et al. 2022's recall-tuned multi-hash pipeline flags 14,310); HyperKvasir and Kvasir-SEG cross-benches against CV2024 return 0 pairs (procedure-disjoint negative controls). At the matched NCC ≥ 0.99 endpoint, the CV2024 rate of 4.66% (540/11,581) is ≈ 148–583× above the ISIC cross-source rate (Wilson-lower / Wilson-upper to point/point).

Controllable 3D part generation is pivotal for digital asset creation, yet existing methods lack the flexibility to accommodate diverse user workflows requiring varying levels of control. In this paper, we introduce FlexPart, a unified image-to-3D part generation framework that supports variable-granularity geometric conditions within a single model. By seamlessly integrating heterogeneous inputs, ranging from sparse points to dense bounding boxes and masks, via gated adaptive modulation, FlexPart unlocks a powerful cross-granularity synergy. We also propose an asymmetric geometric guidance mechanism that leverages strong spatial priors to align representations across granularities, improving geometric consistency and facilitating inter-task synergy. Extensive experiments demonstrate that this synergistic approach outperforms state-of-the-art methods by over 10.5\% on Part CD, achieving high-fidelity part generation with superior geometric consistency across all prompt granularities. Code and models will be made publicly available.

Multimodal large language models (MLLMs) have rapidly advanced machine intelligence by integrating visual perception with powerful linguistic reasoning. However existing visual backbones predominantly encode videos in a rigid per-frame manner and delegate temporal modeling entirely to the language model. This paradigm fundamentally constrains long-range temporal understanding and imposes severe computational inefficiencies. To resolve these bottlenecks we propose FuxViT, a unified Vision Transformer natively supporting a 64K-token context window to seamlessly process diverse inputs ranging from static images to hour-scale videos. By employing a joint spatiotemporal pretraining paradigm, FuxViT generates highly coherent and information-rich visual representations. This strategic early fusion frees the downstream language model from reconstructing low-level temporal structures to focus entirely on high-level reasoning. Furthermore we introduce Frame Selective Attention (FSA) as a native sparse routing mechanism. FSA dynamically attends to the most relevant historical frames to achieve exceptional computational efficiency without sacrificing fine-grained spatial or temporal details. Extensive experiments across 16 image and video benchmarks demonstrate that FuxViT consistently outperforms existing baselines and establishes itself as a vastly superior alternative to standard static ViTs for multimodal understanding. We will fully open-source all model checkpoints.

LLMs are developing increasing capabilities in situational awareness, stealth and autonomy, approaching levels where they may be capable of scheming against their designers by covertly acting on misaligned goals (such as sabotaging oversight mechanisms under cover of a refactoring task). We evaluate model propensity for scheming using honeypots - realistic opportunities to pursue a misaligned goal. We present AutoHoney, an automated pipeline to generate, parameterize, and evaluate scheming honeypots in production codebases. Existing scheming propensity evaluations focus on catching and observing scheming behavior in artificial test settings, whereas we deploy honeypots in real codebases. To study scheming propensity in current models, we use prompted model organisms (PMOs) where goal conflicts are introduced through system prompts, planted documentation, or codebase-embedded environmental cues. Using PMOs with varying levels of goal nudging, we deploy honeypots across four production codebases and conduct a systematic manual ablation on multiple frontier models. Results show that goal presence, threat intensity, and task scope are strong predictors of scheming rates in our model organisms, while urgency has limited effect. The specific goal matters less than its existence—arbitrary goals like "promote mechanical pencils," environmental advocacy, and library funding all produce substantial scheming. Evaluation awareness typically suppresses scheming, but we observe notable exceptions where models take honeypots despite recognizing the evaluative context.

Gaussian Process (GP) kernels are central to Bayesian optimization (BO), yet designing effective kernels for high-dimensional problems still relies on extensive manual engineering. Existing kernel design methods restrict the search to additive and multiplicative compositions of base kernels, while LLM-based BO approaches condition on raw observations, which are infeasible in high dimensions due to context-length constraints and the difficulty of extracting meaningful patterns from high-dimensional numerical observations. We introduce Kernel Discovery, an LLM-driven population-based evolutionary framework that overcomes both limitations. Motivated by the observation that directly prompting an LLM to generate kernel code yields syntactically varied but functionally identical kernels, we adopt a two-stage approach: an LLM first proposes novel mathematical forms, then a second LLM call converts each form into validated, executable code. We also propose a leave-one-out continuous ranked probability score (LOO-CRPS) as a held-out predictive scoring criterion that penalizes overconfident fits more directly than marginal log-likelihood. On five standard high-dimensional BO benchmarks, our method achieves an average rank of 1.2 out of 17, consistently outperforming competitive baselines. We further analyze the discovered kernels to examine which kernel characteristics lead to superior performance in high-dimensional BO.

Multimodal models have demonstrated strong capabilities in web, desktop, and mobile GUI environments, but their performance in automotive cockpit systems remains largely unexplored. In-vehicle GUIs introduce unique challenges beyond traditional GUI interaction, including vehicle-state reasoning, implicit intent understanding, and safety-critical decision making. To address this gap, we introduce Automotive-ENV, an interactive benchmark for evaluating multimodal models in automotive cockpit environments. Built on Android Automotive OS, Automotive-ENV combines cockpit-screen interaction with vehicle-state signals and deterministic state-based evaluation. The benchmark contains 420 carefully constructed tasks covering routine cockpit operation and safety-critical scenarios. Unlike prior GUI benchmarks that rely on action matching or LLM-based judging, Automotive-ENV evaluates model behavior through reproducible validators over UI states, vehicle signals, and safety-related conditions. We benchmark representative general-purpose multimodal models and GUI-specialized vision-language models on Automotive-ENV. Experimental results show that current models remain far from reliable in automotive environments, with substantial performance degradation in implicit and safety-critical tasks. Further analysis reveals consistent failure modes in safety-aware reasoning, vehicle-control understanding, and implicit hazard recognition. We will release Automotive-ENV, together with tasks, validators, and evaluation tools, to support future research on automotive multimodal systems.

A photorealistic and immersive human avatar experience demands capturing fine, person-specific details such as cloth and hair dynamics, subtle facial expressions, and characteristic motion patterns. Achieving this requires large, high-quality datasets, which often introduce ambiguities and spurious correlations when very similar poses correspond to different appearances. Models that fit these details during training can overfit and produce unstable, abrupt appearance changes for novel poses. We propose a 3D Gaussian Splatting avatar model with a spatial MLP backbone that is conditioned on both pose and an appearance latent. The latent is learned during training by an encoder, yielding a compact representation that improves reconstruction quality and helps disambiguate pose-driven renderings. At driving time, our predictor autoregressively infers the latent, producing temporally smooth appearance evolution and improved stability. Overall, our method delivers a robust and practical path to high-fidelity, stable avatar driving.

Aligning multimodal generative models with human preferences demands reward signals that respect the compositional, multi-dimensional structure of human judgment. Prevailing RLHF approaches reduce this structure to scalar or pairwise labels, collapsing nuanced preferences into opaque parametric proxies and exposing vulnerabilities to reward hacking. While recent Rubrics-as-Reward (RaR) methods attempt to recover this structure through explicit criteria, generating rubrics that are simultaneously reliable, scalable, and data-efficient remains an open problem. We introduce Auto-Rubric as Reward (ARR), a framework that reframes reward modeling from implicit weight optimization to explicit, criteria-based decomposition. Before any pairwise comparison, ARR externalizes a VLM’s internalized preference knowledge as prompt-specific rubrics, translating holistic intent into independently verifiable quality dimensions. This conversion of implicit preference structure into inspectable, interpretable constraints substantially suppresses evaluation biases including positional bias, enabling both zero-shot deployment and few-shot conditioning on minimal supervision. To extend these gains into generative training, we propose Rubric Policy Optimization (RPO), which distills ARR’s structured multi-dimensional evaluation into a robust binary reward, replacing opaque scalar regression with rubric-conditioned preference decisions that stabilize policy gradients. On text-to-image generation and image editing benchmarks, ARR-RPO outperforms pairwise reward models and VLM judges, demonstrating that explicitly externalizing implicit preference knowledge into structured rubrics achieves more reliable, data-efficient multimodal alignment, revealing that the bottleneck is the absence of a factorized interface, not a deficit of knowledge.

LLM agents remain unreliable on long-horizon tasks, where small local failures can compound over extended interactions and lead to overall task failure. Although external harnesses can substantially improve robustness, harness design remains a manual and expensive process that requires searching over a large space of prompts, tool configurations, and control logic. We propose AutoSaddler, an automatic harness optimization framework that formulates harness improvement as an offline learning problem and iteratively updates the harness using failure signals from mini-batches. AutoSaddler combines failure-trace diagnosis, structured patch generation that treats the harness as code, and validation-based update selection. Experiments on GAIA2 and Terminal-Bench 2.0 show that AutoSaddler substantially improves agent performance over the corresponding base harnesses, achieving gains of 9.0 and 10.0 percentage points, respectively. Ablation studies further suggest that effective harness optimization benefits from three ingredients: deep debugging rather than shallow reflection, targeted modifications rather than unconstrained editing, and generalization-aware selection rather than trajectory-specific repair. Together, these results suggest that automatic harness optimization is a promising path toward more performant and reliable agent systems.

Scientific research proceeds through iterative cycles of hypothesis generation, experiment design, execution, and revision, often requiring researchers to explore multiple competing directions as evidence accumulates and priorities shift. LLM agents can automate parts of this process, but existing agents either concentrate reasoning within a single research thread or coordinate through a central planner with fixed objectives. As a result, they struggle to sustain parallel exploration across research directions or reorganize as promising and unproductive directions emerge over time. We introduce AutoScientists, a decentralized team of AI agents for long-running computational scientific experimentation. Rather than following decisions from a central orchestrator, agents independently interpret a shared experimental state, self-organize into teams around research directions, critique and filter proposals with a discussion phase before committing experimental compute, and exchange both successful and failed findings across teams to avoid redundant exploration. Under matched experimental budgets, AutoScientists outperforms prior agentic systems across biomedical machine learning, language-model training optimization, and protein fitness prediction. On BioML-Bench, spanning biomedical imaging, protein engineering, single-cell omics, and drug discovery, AutoScientists achieves a mean leaderboard percentile of 74.4% across 24 tasks, improving over the strongest prior biomedical agent by +8.33%. On GPT training optimization, AutoScientists reaches a target validation bits-per-byte 1.9x faster than autoresearch and continues discovering improvements from a stronger starting champion where the single-agent approach finds none (7 vs. 0 accepted improvements). On ProteinGym fitness prediction, AutoScientists discovers a method for ACE2-spike binding that improves over current state-of-the-art model by +12.5% Spearman correlation. Applied without modification to all 217 ProteinGym assays, the same method improves over the prior state of the art by +6.5% in Spearman correlation. Code: https://anonymous.4open.science/r/autoscientists_anonymize-D10C

Geometry Problem Solving (GPS) serves as a rigorous touchstone for evaluating the complex reasoning capabilities of Multimodal Large Language Models (MLLMs), particularly when solutions necessitate auxiliary constructions absent in the visual input. However, current approaches face the following challenges: explicit diagramming methods suffer from unstable image generation quality and error cascading, while existing implicit attention mechanisms are restricted to visible elements, failing to perceive the latent auxiliary structures required for deduction. To bridge this gap, we propose AuxCA (Auxiliary-Clues-Aware GRPO), a novel framework that internalizes the capability of auxiliary construction into the model’s reasoning intuition without relying on external tools. Specifically, we introduce Auxiliary Clues Dependency (ACD) to quantify the causal influence of visual cues and integrate it into GRPO via a fine-grained advantage weighting strategy. Complementing this, we incorporate an auxiliary incentive reward to explicitly encourage the model to actively attend to and utilize these geometric elements during reasoning. Extensive experiments demonstrate that our resulting model, AuxCA-8B, significantly outperforms state-of-the-art explicit plotting methods and achieves performance competitive with top-tier proprietary models. Code is available in \url{https://anonymous.4open.science/r/AuxCA-5510}.

Modern Vision-Language Models (VLMs) benefit from chain-of-thought prompting and test-time scaling, but these gains often come with prohibitive inference cost due to large visual contexts and long decoding chains. We view this cost through two coupled axes: Visual Context Scaling (VCS), which controls how much visual evidence is passed to the language model, and Visual Reasoning Scaling (VRS), which controls how much inference-time reasoning search is performed. Existing methods typically optimize one axis at a time, leaving the joint allocation of compute across these axes underexplored. We introduce Adaptive Visual Inference Scaling (AVIS), a lightweight policy that adapts both VCS and VRS per query. AVIS realizes VCS through Key Diversity Visual (KDV) pruning, a training-free $O(N)$ key-based rule for removing redundant visual tokens before prefilling, and realizes VRS through adaptive self-consistency, using a learned difficulty predictor to select the number of reasoning rollouts. AVIS is deployment-friendly and compatible with shared-prefill inference, where all rollouts reuse a single prefilling pass and KV cache. Across diverse image and video reasoning benchmarks, AVIS improves the accuracy--compute trade-off relative to VCS-only and VRS-only baselines, and remains effective on top of RL post-trained VLMs while keeping compute and latency low.

Open multi-agent systems, where only a subset of agents are controllable, are central to real-world domains such as smart grids, yet lack a principled foundation for reinforcement learning algorithm design. We study this setting through n-agent ad hoc teamwork (NAHT), a recent framework for learning in open multi-agent environments. We propose an axiomatic framework that derives value function learning from the Shapley axioms (Efficiency, Symmetry and Linearity), treating them as structural constraints rather than heuristic design choices. This formulation builds on a novel NAHT decomposition, which characterises the fundamental difference between open and fully controllable multi-agent systems, and establishes a cooperative game-theoretic foundation for learning. We further show that enforcing these axioms on learning individual value functions recovers the Shapley value for Dec-POMDPs, providing a principled bridge between reinforcement learning and classical cooperative game theory. Building on this framework, we introduce Shapley Machines and Banzhaf Machines, two generic algorithmic templates that can be instantiated with standard reinforcement learning methods. When applied to strong baselines such as IPPO and POAM, our methods consistently improve learning efficiency and generalisation to unseen agent types and conventions across NAHT benchmark tasks.

Executable world models can be read, edited, executed, and reused for planning, but only if the program captures the environment's transition law rather than semantic shortcuts in its surface vocabulary. We study online executable world-model learning under prior misalignment, where an agent must induce state-dependent dynamics from interaction evidence alone, without rule descriptions, reward signals, or trustworthy lexical priors. We introduce Alice, a closed-loop system that treats failed candidate updates as structural signal: when a candidate explains a new transition but loses previously explained ones, the preservation conflict reveals dynamics that the current program had conflated. Alice refines these conflicts into hypothesis classes that both provide compact, class-stratified preservation counterexamples for update and guide frontier exploration toward transitions that are novel and underrepresented with respect to the current program. We evaluate Alice on Baba in Wonderland, a prior-misaligned variant of Baba Is You that preserves simulator dynamics while replacing semantically meaningful rule-property labels with unrelated words. Experiments show that Alice substantially improves executable world-model learning under prior misalignment, and ablations show that both class refinement and class-aware exploration contribute.

Contextual bandits require balancing nonlinear reward modeling with online efficiency. Tree ensembles capture nonlinearities but require periodic retraining and large replay buffers. Linear models update efficiently per observation with $O(1)$ memory, but are fundamentally restricted to linear reward structures. We propose Residual Quantization (RQ) as a representation layer to bridge this gap. An offline-trained RQ codebook maps continuous contexts into discrete centroid assignments across $\ell$ levels, set dynamically through a shadow mechanism. This enables a spectrum of additive bandit algorithms that achieve nonlinear expressivity with strictly bounded memory. Across 13 datasets, RQ variants beat their non-RQ counterparts on 11 of 13 datasets, often by wide margins, while matching doubling-retrain XGBoost and neural baselines at up to $1000\times$ less memory.

Long-tailed recognition is commonly formulated as a problem of imbalanced supervision, where rare classes suffer from insufficient training examples. In this work, we study a complementary source of difficulty: under severe scarcity, tail classes may be forced into fine-grained discrimination before the model has acquired sufficiently reliable coarse semantic structure. We refer to this phenomenon as granularity mismatch under scarcity. To address it, we propose BasicLT, a two-stage framework that augments an underlying fine-grained recognition model with basic-level abstraction and commitment-guided selective differentiation. In Stage 1, BasicLT learns an explicit basic-level branch and aligns it with the family-level distribution induced by the fine-grained predictor, encouraging stable coarse semantic organization before additional fine-grained refinement is introduced. In Stage 2, a frozen Stage 1 teacher provides stable family-level commitment signals, and a family-conditioned residual refinement module is trained only on committed medium-shot and few-shot samples. At inference time, the residual correction is activated conservatively according to student-side commitment, so that fine-grained refinement acts as a conditional within-family correction rather than an unconditional auxiliary classifier. We evaluate BasicLT on CIFAR-100-LT, CIFAR-10-LT, ImageNet-LT, iNaturalist 2018, and Places-LT. Across these benchmarks, BasicLT achieves competitive performance against strong long-tailed baselines, with the most consistent gains appearing in medium-shot and few-shot regimes. Ablations and analyses further show that both basic-level abstraction and commitment-guided selective refinement are important, supporting the view that controlling the semantic granularity of discrimination can complement conventional rebalancing strategies for long-tailed learning. Code is available at Supplement.

Large language models and their vision counterparts have achieved remarkable capabilities, yet their wider deployment is increasingly constrained by the computational and memory demands of Transformer architectures. Much recent work improves Transformer efficiency, but the query-key pathway is still typically parameterized by fully independent per-head projections, which place a substantial parameter burden on attention. Motivated by the view that token selection may admit a shared low-dimensional geometry, we propose $\textbf{Basis-Mediated Bilinear Attention} (\textbf{BMB})$, a training-time reparameterization of the query-key pathway in which all heads within a layer share a latent basis while retaining head-specific interactions. A practical explicit variant, BMB-UV, materializes per-head query and key tensors through a shared basis and lightweight head-specific factors, reducing the query-key parameter count from $2d^2$ to $dr+2Hrs$. We evaluate our method family on ViT-Base image classification on ImageNet-1K and on GPT-2 small pretraining on the C4 RealNewsLike subset, and compare it against recent query-key parameter-sharing and low-rank baselines. Relative to standard multi-head attention, our method family reduces query-key parameters by up to $ \textbf{95.83} $%$ $ while remaining competitive with the baseline on both benchmarks.

Feed-forward transformer models have driven remarkable progress in 3D vision, yet their quadratic complexity renders them impractical for long video sequences. Streaming 3D reconstruction addresses this bottleneck by processing frames sequentially with constant memory. However, existing recurrent architectures inevitably suffer from progressive degradation over long sequences, as they fail to account for the intrinsic reliability of observations and the accumulated confidence of the historical state, leading to severe error accumulation. To address these issues, we introduce BAT3R, a training-free Bayesian adaptation framework comprising two synergistic modules: State Innovation Gating (SIG) and Bayesian-guided Adaptive State Updating (BASU). SIG acts as a distributional regularizer that anchors state evolution to its geometric initialization, effectively arresting representational drift during long-term inference. Building on this, BASU reformulates state update as a recursive Bayesian inference process within a Kalman filtering framework. By jointly modeling the temporal reliability of observations via information-theoretic gain and spatial relevance via cross-attention maps, BASU achieves an asymmetric spatio-temporal state fusion: well-consolidated historical regions are protected from redundant updates, while genuinely novel scene content is aggressively incorporated. Extensive experiments on standard benchmarks (7-Scenes, TUM, ScanNet, KITTI, etc.) demonstrate that BAT3R consistently outperforms state-of-the-art online 3D reconstruction methods, with the most significant gains observed on long sequences and complex scenes.

Bayesian Optimization has become an established methodology for minimizing black-box functions of a vector input. Often, however, this parameter vector arises from the discretization of an inherently functional relationship. Several recent articles have considered the Functional Bayesian Optimization (FBO) setting, in which the variable to be optimized is not a member of a finite dimensional vector space, but rather an infinite dimensional function space. In this work, we propose $L^0$ Manifold Optimization (L0MO), a simple approach to FBO which searches a sub-manifold of a Reproducing Kernel Hilbert Space consisting of functions with a sparse representation in the kernel functions. This results in simpler, more effective functional solutions than existing methods. We discuss in detail the relationship between our method and existing ones, providing a unifying lens through which to view prior works. To assess our method against the state of the art, we conduct an extensive computational study, and along the way develop a novel set of benchmark test functions which port standard finite-dimensional ones to the infinite dimensional domain. Our experiments demonstrate that the proposed method achieves superior performance across a wide range of test benchmarks.

AI evaluation pipelines often produce conflicting judgments rather than clean labels. In pairwise LLM evaluation, this conflict is especially visible: disagreement can arise from ambiguous items, underspecified rubrics, heterogeneous or unstable human raters, or an LLM judge whose verdict changes when the response order is swapped. We propose BayesJudge, an online Bayesian meta-evaluation layer for conflicting human-LLM judgment streams. For each comparison, BayesJudge estimates a panel-relative posterior verdict distribution over the two responses, with a tie or ambiguity state when such labels are available. At the same time, it estimates rater-specific human confusion matrices and LLM presentation-order bias. The method uses tie-open labels to keep ambiguity observable and paired order-swapped judge calls to separate response quality from presentation effects. We formulate the exact online posterior recursion and use a scalable Rao--Blackwellized assumed-density SMC approximation for streaming inference. Controlled synthetic experiments demonstrate recovery of prespecified evaluator parameters and illustrate two protocol-level identifiability mechanisms: tie-open labels expose ambiguity mass, and order-swapped paired judgments separate item preference from position bias. On real-world SummEval dataset, BayesJudge successfully detects systematic presentation-order effects in LLM judge outputs, infers distinct expert and crowdworker behavior signatures without rater metadata, and produces posterior uncertainty estimates that correlate with human disagreement. Our code is available at https://anonymous.4open.science/r/BayesJudge-0879.

Simulating student learning behaviors in open-ended problem-solving environments holds potential for education research, from training adaptive tutoring systems to stress-testing pedagogical interventions. However, collecting authentic data is challenging due to privacy concerns and the high cost of longitudinal studies. While Large Language Models (LLMs) offer a promising path to student simulation, they suffer from competency bias, optimizing for efficient correctness rather than the erratic, iterative struggle characteristic of novice learners. We present BEAGLE, a neuro-symbolic framework that addresses this bias by incorporating Self-Regulated Learning (SRL) theory into a novel architecture. BEAGLE integrates three key technical innovations: (1) a semi-Markov model that governs the timing and transitions of cognitive behaviors and metacognitive behaviors; (2) Bayesian Knowledge Tracing with explicit flaw injection to enforce realistic knowledge gaps and "unknown unknowns"; and (3) a decoupled agent design that separates high-level strategy use from code generation actions to prevent the model from silently correcting its own intentional errors. In evaluations on Python programming tasks, BEAGLE significantly outperforms state-of-the-art baselines in reproducing authentic trajectories. In a human Turing test, participants could not reliably tell BEAGLE traces apart from real student data: classification accuracy was statistically equivalent to chance (52.8\%, $d'=0.15$, $N=71$).

We propose an instantiation of flow matching that relies on a time-independent velocity field (an \emph{autonomous flow}) to exactly map between two distributions, so long as the target is singular, i.e.\ supported on a lower-dimensional data manifold. We also show that the one-step generative map associated with this flow is the unique solution of a simple conservation equation, which can be used to learn the map directly from samples. These autonomous flows and maps give a dynamical meaning to the flux constraint of Beckmann's transportation problem. Their construction provides a unifying framework that recovers, for instance, the closed-form Poisson-flow generative model and equilibrium matching with a quadratic flow-matching regression loss. We illustrate how this theory corrects inconsistencies in existing methods and improves their performance on large-scale image generation.

Video Large Language Models (VideoLLMs) receive frames in sequential order and interpret how visual content evolves along the temporal axis, yet temporal reasoning remains a persistent weakness across architectures. Reversing the frame order of a video, a transformation that should invert temporal answers, often leaves the final prediction unchanged. We investigate where this failure originates by defining the temporal divergence vector $\tau_l$, the layer-wise representational difference induced by reversing temporal order. Tracking its magnitude across layers reveals a consistent temporal divergence profile where the divergence peaks at intermediate layers and progressively diminishes toward the output. We confirm this peak is specific to temporal reasoning and functionally critical for predictions, establishing that VideoLLMs acquire temporal information at intermediate layers but fail to maintain it to the output. This progressive fading motivates our method, Temporal Activation Injection (TAI), which extracts $\tau_l$ at the peak of the profile for each input and reinjects it into subsequent layers following the measured decay. TAI requires no training and consistently improves temporal reasoning across three VideoLLMs and four benchmarks with negligible impact on non-temporal tasks.

Current video foundation models, including the strongest self-supervised models such as V-JEPA2, fail to capture how humans organize social information in dynamic scenes. For example, across a range of diverse vision models tested, none were able to predict human similarity judgments to social video clips as well as a sentence embedding model of the caption text (MPNet). We show this gap in vision model performance can be closed by a compact behavioral supervisory signal. We introduce behavioral geometric supervision (BGS): a hybrid objective that constrains local and global pairwise embedding geometry to match the relational similarity structure across videos. We apply this method using a new human similarity dataset, containing 49,484 odd-one-out judgments from 250 naturalistic social video clips, and low-rank adaptation across four ViT backbones (V-JEPA 2/2.1, TimeSformer, VideoMAE, and CLIP). We find that one of the best fine-tuned models, V-JEPA 2.1, nearly triples in performance compared to the pre-trained baseline and reaches close to the noise ceiling, exceeding the strongest sentence-embedding baseline. In addition, finetuned models (i) capture unique variance in human judgments that caption-based language embeddings do not, (ii) develop interpretable social-affective attributes (valence, arousal, and dominance) despite never being trained on any of these attributes, (iii) zero-shot transfer to a separate dataset of out-of-distribution abstract social interactions, and (iv) shift spatial attention from scene context to socially informative regions (faces, gaze, and interacting bodies). A matched language-distillation control fails to reproduce these gains, ruling out caption transfer as the mechanism. Our results show how a modest amount of human behavioral data can steer video models toward human-like social visual understanding.

LLM supervision systems, namely input/output moderation filters and jailbreak detectors, are the primary safeguard against misuse in deployed AI applications, yet existing benchmarks are often vendor-biased, omit cost and latency, and rarely compare specialized guardrails against repurposed generalist LLMs. We present BELLS-O (Benchmark for the Evaluation of LLM Supervision Systems -- Operational), the first independent operational benchmark of LLM supervision systems. BELLS-O evaluates 28 systems from 17 providers: every major specialized guardrail (e.g., LlamaGuard-4, ShieldGemma-2, Lakera Guard) and frontier generalists repurposed as supervisors (e.g., GPT-5.4, Claude Sonnet 4.6, Grok-4.1), jointly on detection rate, false-positive rate, latency, and monetary cost. We cover input/output moderation across 11 harm categories and jailbreak detection across 13 attack techniques, using in-house datasets built from handcrafted prompts, expert-curated samples, and quality-controlled synthetic generation. To prevent latent generator-specific signals in synthetic data, every generated sample is run through a paraphrasing step that suppresses these fingerprints. Mapping the Pareto frontier reveals use-case-dependent tradeoffs. On content moderation, specialized supervisors are operationally dominant: top systems match frontier LLMs on detection ($\approx$95\% vs.\ 94\%) at comparably low false-positive rates ($\leq$2\%), while running 5--10$\times$ faster and ${\sim}$10$\times$ cheaper. On jailbreak detection, the tradeoff shifts: frontier LLMs achieve higher detection and lower false-positive rates but at 10--50$\times$ higher cost and 5--10$\times$ higher latency. We release the benchmark, framework, leaderboard, and datasets as the first vendor-neutral basis for selecting safeguards under real deployment constraints.

Machine Learning Interatomic Potentials play a fundamental role in computational chemistry and materials science, enabling applications from molecular dynamics simulations to drug design and materials discovery. While recent approaches can estimate inter-atomic forces with high precision, it remains unclear to what extent they can generalise to previously unseen molecules. Do they learn the compositional structure of chemistry, capturing how molecular fragments and their combinations determine properties, or do they primarily learn to interpolate patterns that are specific to the training examples? To address this question, we propose a benchmark consisting of four tasks that require some form of compositional generalisation. In each task, models are tested on molecules that were unseen during training, but the training data is chosen such that generalisation to the test examples should be feasible for models that learn the underlying physical principles. Our empirical analysis shows that the considered tasks are highly challenging for state-of-the-art models, with errors on out-of-distribution examples often an order of magnitude higher than on in-distribution examples, even when using foundation models that have been pre-trained on millions of molecules.

As language models are increasingly deployed as autonomous agents, they will need to coordinate with others in long-horizon, open-ended interactive tasks. Yet current evaluations rarely test these demands together, focusing instead on short interactions, single-agent open-ended tasks, or highly structured multi-agent settings. We introduce $alem$, a JAX-based, procedurally generated open-ended benchmark for multi-agent coordination built on Craftax-like dynamics. Alem embeds procedurally generated coordination tasks, soft specialisation, communication, and controllable coordination difficulty into a long-horizon survival world with exploration, crafting, trading, and combat. We evaluate $13$ modern LLMs zero-shot within homogeneous teams, with trained MARL agents as reference points. Most LLM agents struggle, averaging only ~6% of maximum reward, although performance varies widely. Gemini-3.1-Pro-High approaches MARL agents trained for one billion steps on Hard coordination ($17.5\%$ vs. $17.6\%$ Coord.\%), Gemma-4-31B-it is the strongest tested open-weight model, and GPT-5.4-High makes strong base progress, while achieving much lower coordination reward. We find that base task competence does not imply coordination competence, communication helps agents share intent, memory and reasoning help with multi-step planning, and heterogeneous teams regress toward average member performance rather than matching their strongest teammate. These results identify coordination as a distinct bottleneck for current LLM agents, separate from single-agent capabilities. Alem makes this bottleneck measurable, providing a controlled setting for developing agents that can communicate, allocate roles, and execute shared plans.

Sequential testing avoids some of the many controversies and drawbacks related to traditional batch testing with p-values. In particular, the testing by betting paradigm has developed rapidly in recent years and frames the level of evidence against the null as the wealth of a player in a betting game. Naturally, how much a player should bet is a key consideration, and should reflect the statistician's confidence in what they know about the next observed data point. Much of current work uses standard betting strategies developed for adversarial problems, where the data can be hand-picked to the detriment of the player. In the stochastic setting, where the data observed is independent and stationary, these approaches may be suboptimal. We propose a simple betting strategy, **follow the leave-one-out e-power (FLOE), that exploits the i.i.d. nature of a data stream to estimate the e-power of our current test on the underlying distribution. We average not over the historical evidence collected so far, as generic online learning-based methods do, but instead over the equally-likely possible data sequences we might have seen based on shuffling the data. With an additional principled regularization technique, this simple drop-in method yields substantial improvements in testing performance on a variety of kernel two-sample testing problems.

Language models (LMs) obtain much of their capability from the diversity of their training data. This is possible because neural architectures like transformers are remarkably flexible and can handle a variety of domains such as natural language, code, and mathematics. This raises the question: are transformers optimal for any specific domain? In this position paper, we present empirical evidence that the inductive biases of transformer-based LLMs should be re-examined, since, while broadly useful, they are often suboptimal on subsets of their training data. Methods. To support this position, we use a method that learns new activation functions within a transformer as a tool to explore the space of inductive biases. We train the resulting modified architectures on various data subsets. A large-scale evaluation shows that vanilla transformers are relatively easy to improve, but mostly on a per-domain basis. For example, an architecture optimized for a natural language dataset like FineWeb performs well on another such dataset like TinyStories, whereas architectures optimized for mathematics perform poorly on natural language. We also identify the mechanisms responsible for these improvements. Implications. These results call for re-evaluating core design choices in language models. We do not advocate a return to handcrafted, domain-specific models. Instead, we emphasize that different domains can benefit from different learning mechanisms. This will require new designs that preserve cross-domain interactions, analogous to how reasoning in the brain arises from the interaction of multiple specialized networks. We argue that this perspective may be key to improving reliability and data efficiency.

Sharpness-Aware Minimization (SAM) is widely used in domain generalization to promote flat minima that generalize to unseen distributions. Existing SAM-based methods typically penalize the sharpness of the *average* loss across domains. We show that this prevailing formulation—termed AvSAM—has a fundamental geometric limitation: it can favor solutions with misaligned domain-wise dominant Hessian eigenvectors. As a result, such solutions may appear flat on average while remaining sharp within individual domains. To address this limitation, we propose an alternative objective that penalizes the average *domain-wise* sharpness. We show that this objective decomposes into the AvSAM objective plus a non-negative residual term, which we call the *Cancellation Gap*. This gap is minimized if and only if the domain-wise dominant eigenvectors are aligned up to sign, indicating that it mitigates AvSAM's tendency to favor misaligned solutions. However, we further uncover a degeneracy: the Cancellation Gap can become small without improving alignment when sharpness is highly imbalanced across domains. To mitigate this failure mode, we propose Cancellation Gap Minimization (CGM), which augments the objective with a squared coefficient-of-variation ($\mathrm{CV}^2$) regularizer to discourage sharpness imbalance. Experiments show that CGM achieves the best average performance across standard domain generalization benchmarks among strong baselines. Diagnostic analyses further confirm reduced domain-wise sharpness and improved alignment of dominant Hessian eigenvectors.

Low-Rank Adaptation (LoRA) and its weight-decomposed variants dominate parameter-efficient fine-tuning, yet their directional update mechanisms remain geometrically mismatched to the normalized parameterization they impose. In existing decoupled PEFT methods, low-rank directional updates are computed in unconstrained Euclidean space and only then projected back onto a spherical manifold, which can introduce capacity redundancy and directional gradient distortion. To address these limitations, we propose Geodesic Orthogonal Low-Rank Adaptation (GO-LoRA), a geometry-aware reparameterization framework for decoupled PEFT. GO-LoRA first applies orthogonal tangent projection to remove components parallel to the pretrained base direction, ensuring that the low-rank directional budget is devoted entirely to tangent-space motion. It then uses a geodesic flow based on the Riemannian Exponential Map to construct manifold-consistent directional updates on the hypersphere, improving geometric fidelity relative to Euclidean normalization while retaining standard optimization over the underlying low-rank parameters. Extensive experiments across language, vision, code generation, and multimodal reasoning tasks show that GO-LoRA consistently outperforms strong LoRA baselines while introducing no additional inference-time overhead after offline merging. The code is available at https://anonymous.4open.science/r/go-lora.

Recent work has shown that contrastive representation learning can be understood as estimating a positive-pair, PMI-like kernel, and that spectral factorization of this kernel yields useful eigenfunction representations. This paper asks what happens beyond squared-error spectral geometry. Modern representation learning objectives are rarely pure $\ell_2$ kernel-approximation objectives: they use conditional KL, InfoNCE, JS/NCE, Brier, Hellinger, least-squares ratio fitting, and other statistical discrepancies. We show that these discrepancies do not merely provide alternative estimators of the same affinity; they induce different local geometries for the extraction of low-rank representations. To formalize this, we introduce divergence principal functions, a geometry-aware generalization of eigenfunctions. Eigenfunctions are recovered as the special case corresponding to squared-error geometry. For smooth discrepancies, we show that divergence principal functions locally solve a weighted spectral approximation problem, with weights given by the curvature of the divergence at the target affinity. This provides a loss-consistent bridge from positive-pair learning objectives to representation geometry and clarifies when spectral/eigenfunction representations are appropriate and when they are mismatched to the training objective.

Monte-Carlo Tree Search (MCTS) typically stores only scalar mean estimates, which makes it poorly suited for tail-risk objectives such as conditional value-at-risk (CVaR). We introduce **Categorical Thompson Sampling with Optimism (CATSO)** and **Particle Thompson Sampling with Optimism (PATSO)**, two distributional MCTS algorithms that maintain finite-support empirical backup laws at Q-edges. CATSO represents Q-edge laws with categorical atoms and Dirichlet Thompson sampling, while PATSO represents them with adaptive particles. Both algorithms select actions using the CVaR of a Thompson-sampled Q-edge law plus a polynomial optimism bonus, and propagate scalar continuation values through visit-weighted averages of child Q-edge CVaR scores. For the finite-depth nested CVaR planning target, we prove an $O(n^{-1/2})$ bound on the root value error after $n$ rollouts, which upper-bounds the risk-sensitive simple regret. CATSO adds a fixed-grid discretization bias, while capped PATSO adds a tunable Wasserstein compression bias. Experiments on stochastic planning benchmarks show that distributional Q-edges improve lower-tail return and reduce catastrophic outcomes when mean-optimal and risk-sensitive behavior differ.

As collaborative machine learning scales to thousands of edge devices, two failure modes emerge: centralized federated learning creates server bottlenecks and single points of failure, while flat peer-to-peer systems force every node to increase its communication degree as the network grows. We propose Tiered Gossip Learning (TGL), a two-layer push–gossip–pull protocol that decouples data-holding leaf nodes from a decentralized relay backbone. In each round, leaves push local models to a subset of relays, relays gossip among themselves to mix models globally, and leaves pull updates from another random relay subset. This asymmetric design keeps per-leaf communication fixed as the network scales, offloading the mixing burden to a thin layer of higher-capacity relays without relying on any central coordinator for aggregation. Across CIFAR-10, FEMNIST, and AG News, TGL matches or exceeds baseline accuracy with up to 80\% fewer model exchanges. We provide convergence guarantees under standard smoothness, bounded variance, and heterogeneity assumptions, with explicit stage-wise consensus bounds characterizing how relay-layer connectivity governs the global mixing rate.

Foundation models for predictive machine learning on tabular data have recently gained significant traction in academia and industry. Research communities across disciplines are increasingly evaluating tabular foundation models on diverse datasets and tasks. However, these task- and discipline-specific evaluations remain largely inaccessible to model researchers because benchmark software and evaluation protocols are fragmented. As a result, model researchers rely on standard benchmarks, which are mostly defined for tasks where tabular foundation models already excel. The most challenging scenarios are excluded, limiting meaningful progress in the field by focusing on marginal improvements on IID data rather than on broader, more demanding challenges. To overcome this, we introduce BeyondArena, the first unified holistic benchmark for tabular data that supports diverse task types (IID, temporal, grouped), across sample size and feature dimensionality scales, with diverse feature types (with text, with high cardinality) from a broad range of disciplines. To enable unified benchmarking beyond standard benchmarks, we introduce DataFoundry, a Python framework and metadata schema for curating tabular datasets for predictive machine learning. Our results across 11 models and 142 curated datasets show that existing tabular foundation models excel on tiny- to medium-sized IID data, while traditional tree-based and deep learning models still dominate on non-IID, large, and high-dimensional datasets. BeyondArena guides model research for the most demanding challenges in tabular data, enabling progress towards truly foundational tabular models.

We introduce a new algorithm to sample from a probability density $\pi \propto e^{-V}$ with $V(x): \mathbb{R}^d \rightarrow \mathbb{R}$ a strongly multi-modal potential. Such distributions are challenging to sample from using standard gradient based methods, as exploration typically relies on inefficient diffusions. Inspired by connections between Schrodinger operators, Langevin diffusions and Morse theory, we develop a particle based method that exploits fast relaxation to transition pathways of the so-called Witten PDE on 1-forms (vector fields). Once pathways joining modes are sufficiently well explored, the method relaxes to inexpensive Langevin dynamics for sampling within approximately convex basins. Numerical experiments demonstrate superior performance over cheap gradient based sampling methods and competitive performance at lower cost compared to a state of the art method, parallel tempering.

Neural denoisers trained with pointwise distortion losses such as MSE often suppress the temporal regularities that downstream tasks depend on---a failure that becomes acute under non-Gaussian, impulsive, or non-stationary disturbances. We propose a training framework that lifts a family of classical complexity descriptors---permutation entropy, sample entropy, Lempel-Ziv complexity, and Higuchi fractal dimension---into end-to-end differentiable objectives via calibrated soft surrogates with straight-through gradients. Three complementary mechanisms emerge from this lifting: (i) early fusion injects local structural priors as auxiliary inputs, (ii) a complexity-consistency regularizer matches the multi-scale complexity of denoiser outputs to clean references, and (iii) a curriculum sampler prioritizes examples whose noise-induced complexity inflation is largest. We prove that additive noise strictly inflates ordinal uncertainty for bounded piecewise-smooth signals, and that, under bounded surrogate calibration, complexity consistency yields a strictly larger expected decision margin than MSE-only training. On OFDM/QAM communication signals under AWGN, $1/f$ colored, and $\alpha$-stable impulsive noise, our framework reduces post-detection bit error rate by 29--47% over Transformer (PatchTST, iTransformer), GNN, and CNN baselines, with statistically significant gains across five seeds and 95% bootstrap confidence intervals. The improvements transfer to over-the-air conditions on the public RadioML 2018.01A benchmark (single-carrier QAM portion) and a new 50-session in-house SDR testbed, yielding 22--26% relative BER reduction without fine-tuning. The framework is architecture-agnostic and adds $\leq 3.5\%$ FLOPs over the underlying backbone.

Graph foundation models have emerged as a promising paradigm for generalizing across diverse structural tasks. Current approaches serialize graphs into node token sequences and apply autoregressive or masked prediction, yet this node sequential modeling disrupts the relation-centric nature of graphs and violates permutation equivariance. We propose relational diffusion, a framework that reconceptualizes graph modeling through edge co-occurrence modeling. By treating edges as fundamental modeling units and learning their joint distribution via discrete diffusion processes, our approach naturally respects the relational semantics of graphs while maintaining permutation equivariance. We introduce topological edge tokenization to enrich edge representations with multi-hop structural context, and relational prompt completion to unify node classification, link prediction, and graph classification as sequence completion problems over edge representations. Through extensive experiments, we demonstrate that relational diffusion achieves strong performance while providing a principled foundation aligned with the relational nature of graphs.

This paper considers a practical few-shot anomaly detection (FSAD) setting, termed discriminative FSAD, where a limited number of both normal and anomalous examples are available as references during inference. Existing FSAD methods rely on normal-only references through normality matching, ignoring the discriminative clues in anomalous references, while directly fitting both references can overfit to the seen anomalies. We introduce IDEAL, an intrinsic deviation learning framework that leverages both reference types to learn intrinsic deviation patterns characterizing generalizable abnormality as deviations from normality. IDEAL decomposes the learning process into two novel components: 1) a Normal Variation Eraser to suppress nuisance normal variations that may lead to noisy deviations from normality, thereby highlighting anomaly-relevant deviation representations; 2) an Intrinsic Deviation Encoder to decompose these denoised deviation representations into intrinsic deviation vectors capturing the most discriminative orthogonal deviation directions. At inference, IDEAL scores query-to-normal deviations preserved after projection onto the learned intrinsic deviation vectors, enabling generalization for both seen and unseen anomalies. Extensive experiments on eight real-world datasets show that IDEAL generalizes effectively to unseen anomalies and consistently outperforms existing state-of-the-art FSAD methods.

Automatic Prompt Optimization (APO) aims to improve the capabilities of Large Language Models (LLMs), yet existing methods remain limited by their result-oriented nature. Training-free approaches often converge to suboptimal prompts, while training-based methods suffer from reward sparsity and optimization instability. More importantly, both paradigms focus on obtaining the final prompt rather than learning the prompt optimization process itself. To address this dilemma, we propose \textbf{ProPO}, a trajectory-driven \textbf{Pro}cess learning framework for \textbf{P}rompt \textbf{O}ptimization guided by multi-dimensional composite rewards. Specifically, ProPO introduces a historical trajectory reflection mechanism for high-quality exploration. Concurrently, we adopt adaptive rubric rewards to provide fine-grained supervision and alleviate reward sparsity. More importantly, we design a novel Process Logic Reward that supervises the consistency of the reflection process. Extensive experiments demonstrate that ProPO, integrated with GRPO, establishes a new SOTA performance. Notably, ProPO achieves a remarkable \textbf{70.00\%} accuracy on the highly complex AIME-2026 task and an absolute gain of \textbf{+25.25\%} over the zero-shot baseline on Subjectivity Classification (Subj). Furthermore, empirical analyses reveal that our multi-dimensional reward mechanism provides finer-grained and denser signals, allowing the model to learn the prompt optimization process, and ultimately enabling it to generate significantly higher-quality prompts.

Video Reasoning Segmentation (VRS) demands both complex semantic reasoning and robust spatiotemporal mask propagation. Existing paradigms typically compress reasoning into an isolated, opaque token (e.g., \texttt{[SEG]}) via expensive fine-tuning, or rely on training-free prompt-then-track pipelines that inevitably suffer from semantic degradation and identity switches when encountering hard distractors. To fundamentally address these vulnerabilities, we propose Counterfactual Logic and Explicit Anchoring Reasoning (\textbf{CLEAR}) framework, a training-free VRS framework inspired by human dual-system cognition. Specifically, we introduce a Causal-Counterfactual Inference mechanism that transforms error-prone coordinate regression into a tractable discrete instance selection task, establishing highly reliable initial anchors via rigorous bidirectional reasoning. During mask propagation, we design an adaptive mode-switching mechanism bridged by a Heterogeneous Event Monitor. The fast-thinking mode utilizes a purely visual engine for efficient temporal propagation, while the Monitor continuously evaluates its reliability across multiple dimensions. Upon detecting propagation anomalies, the slow-thinking mode intervenes and explicitly replaces \texttt{[SEG]} tokens with a structured Contrastive Concept Dictionary for closed-loop identity recovery..Extensive experiments demonstrate that CLEAR significantly outperforms existing training-free methods and achieves performance comparable to state-of-the-art training-based approaches on comprehensive referring and reasoning VOS benchmarks. Code will be available after acceptance.

Semi-supervised medical image segmentation has gained increasing attention for its outstanding performance with limited annotations. Most methods follow the Mean Teacher framework, where Exponential Moving Average (EMA) is primarily used to stabilize predictions, while its potential roles beyond stabilization are largely overlooked. In this work, we interpret EMA as a first-order infinite impulse response (IIR) low-pass filter with inherent frequency selectivity, which enables the preservation of semantic information at different levels. Motivated by this insight, we propose a dual-EMA teachers framework, including a long-term teacher focusing on global structures and a short-term teacher emphasizing local details. Furthermore, to help the student better exploit the guidance from both teachers, a First-In-First-Out replay queue is designed to reuse historical high-confidence pseudo-labeled pairs, leveraging pseudo labels provided at different training stages and reducing the sensitivity to sampling bias. Meanwhile, we impose feature-level global semantic and local relation consistency constraints to alleviate the impact of noisy pseudo labels, facilitating robust transfer of complementary representations from dual teachers to the student model. Experiments on three public medical image datasets demonstrate that our method outperforms current SoTA methods, verifying its effectiveness.

Reinforcement learning (RL) has emerged as a paradigm for fine-tuning large-scale generative models, such as diffusion and flow models, to align with complex human preferences and user-specified tasks. A fundamental limitation remains *the curse of diversity collapse*, where the objective formulation and optimization landscape inherently collapse the policy to a Dirac delta distribution. To address this challenge, we propose **DRIFT** (**D**ive**R**sity-**I**ncentivized Reinforcement **F**ine-**T**uning for Versatile Image Generation), an innovative framework that systematically incentivizes output diversity throughout the on-policy fine-tuning process, reconciling strong task alignment with high generation diversity to enhance versatility essential for applications that demand diverse candidate generations. We approach the problem across three representative perspectives: i) **sampling** a reward-concentrated subset that filters out reward outliers to prevent premature collapse; ii) **prompting** with stochastic variations to expand the conditioning space, and iii) **optimization** of the intra-group diversity with a potential-based reward shaping mechanism. Experimental results show that DRIFT exhibits clear Pareto dominance in task alignment and generation diversity, achieving 7.19\%$\sim$93.40\% higher diversity at matched alignment and 13.23\%$\sim$60.13\% higher alignment at matched diversity.

We introduce a constructive universality principle for structural interpretability in binary regression. Let $A\in \\{-1,1\\}^{p}$ be binary covariates and $B\in\\{-1,1\\}$ be a binary response. We represent the binary-regression target $\mathbb{E}[B|A]$ by Ising Hamiltonians over observed and auxiliary binary variables. This construction encodes covariate configurations as Boolean states and maps input-output relationships through faithful reductions from Boolean satisfiability (SAT) to ground-state-energy Ising problems. This gives a interpretable graph representation of high-order binary dependence. We showcase this principle with the *Two-body, Linear, Universal, Binary Effect* (T-LUBE) framework. Unlike full power-vector expansion, T-LUBE does not enumerate high-order product features. It recasts these effects through interpretable auxiliary variables and pairwise interactions on an augmented binary graph. Proof-of-concept experiments corroborate our theory.

Learning meaningful representations from cellular microscopy images remains a central challenge in computational biology and computer vision. Recent advances in self-supervised learning (SSL) have shown that scaling model and dataset size can improve the recovery of biological relationships such as gene–gene or compound–target associations. However, this progress has been driven primarily by scale rather than structure—requiring vast data and compute while offering no guarantee that the resulting embeddings capture underlying biology. We introduce NOVA, a structure-first framework for learning biologically aligned representations of cellular perturbations. NOVA integrates curated biological graphs directly into the training objective of a vector-quantised transformer, aligning embeddings of related genes and pathways through graph-based regularisation. By incorporating these relational priors during training, NOVA learns biologically faithful representations with orders of magnitude less data and compute than current state-of-the-art SSL models, while achieving superior performance on gene-gene retrieval benchmarks. This approach reframes representation learning for microscopy from a problem of scale to one of structure, offering a more systematic and efficient route toward biologically grounded embeddings.

Biomicroscopy imaging involves the complex coordination of multiple procedures, such as focusing, illumination, and staining. Traditional unitary methods struggle with dynamical environments, while hybrid models pre-trained on natural images tend to introduce artifacts for higher perceptual metrics. Notably, biological imaging is not a one-off task. It is an iterative process similar to human expert’s observation, verification and adjustment. They can benefit from history, remembering which processing strategies are suitable and can also identify issues such as artifacts and mis-staining from verification. To this end, we propose BioMicroAgents, the first multi-agent framework with co-evolutionary capabilities designed to simulate and optimize whole biomicro imaging process including brightening, denoising, deblur, super-resolution, and virtual staining. Through reinforcement learning, the model can reflect based on historical experience to reduce artifacts and mis-staining, while enhancing edge and structural features to improve the accuracy of downstream tasks. Specifically, BioMicroAgents includes: (i) Multi-modal Analytical Agents, which utilize multimodal RAG to provide template images and biomicro knowledge, and employ the memory mechanism to analyze optimal processing paths; (ii) We propose three biomicro fidelity metrics and utilize three agent groups (MicroNavigator, MicroProcessor, and MicroAuthenticator) to perform rigorous analysis, processing, and verification. (iii) Co-evolutionary ability, which achieves dynamic policy and self-correction through interactive feedback strategies and memory mechanisms, allowing the agent to evolve from history and other agents. We test model across 12 biomicro tasks, BioMicroAgents achieved improvements in perceptual quality while adhering to strict fidelity in sub-cellular features. It also enhances the accuracy of downstream tasks such as segmentation and cell counting, showcasing the potential of general-purpose visual agents in bioscience.

Biological risk is often a property of a cascade rather than a single model output. A generated artifact may become consequential only after transcription, translation, folding, binding, pathway propagation, or host-directed off-target interaction. Existing bio-AI safety evaluations largely grade each model within its own modality, missing downstream risks and inheriting self-grading circularity when model and evaluator share representations. We introduce BioSafetyBench, an agentic cascade-evaluation platform that treats the bio-AI ecosystem as the unit of safety evaluation. BioSafetyBench routes each model-output pair to one of 15 standardized probes over a six-level hierarchy from genome to organism-level outcome. Fixed bridge agents propagate outputs through two causal pathways: Pipeline A follows forward biological cascades, while Pipeline B evaluates CRISPR and siRNA designs through genome- and transcriptome-wide off-target alignment. At every reached level, external frozen evaluators compute domain-native biomarkers and compare them with literature-anchored thresholds, yielding a Cascade Concern Profile and Cascade Depth. Across 25 foundation models, BioSafetyBench surfaces risks invisible to single-modality grading: Influenza neuraminidase designs from ProteinMPNN and ESM-IF1 reach CD = 4; jailbreak prompts inflate gRNA off-target hits 9.05× over baseline, push 48 of 604 gRNAs to Critical, and concentrate Critical siRNA hits on MYC; protein mask-and-fill spans a 32× cross-model gap; GPT-4o complies on all natural-language-guided protein mutation prompts at CD = 2 while Claude Sonnet 4.5 refuses all prompts at CD = 0; and a three-predictor ADMET ensemble misses 30.3% of known clinical toxics. BioSafetyBench is, to our knowledge, the first bio-AI safety benchmark to provide cascade-level risk certificates whose scoring modules are weight- and gradient-flow independent from the model under test by construction.

Post-training pruning of large language models (LLMs) zeroes a fraction of the weights and reconstructs the survivors without retraining, addressing the weight-memory bottleneck of LLM serving. At a fixed mask the per-row reconstruction admits a closed-form Optimal Brain Surgeon (OBS) optimum, but the cost of the exact joint solve is cubic in the layer dimension; one-pass methods such as SparseGPT and Wanda approximate it at quadratic cost without attaining the per-block joint optimum. We introduce Block-OBS-GS, which solves the per-block joint OBS reconstruction exactly within SparseGPT's complexity class via a per-row Cholesky factorisation of the in-block damped Hessian (one third the cost of an explicit inverse, reused across the algorithm) and follows it with a single cross-block Gauss--Seidel sweep. We prove three guarantees: per-layer FLOP count in SparseGPT's complexity class, with the leading $mn^{2}$ constant dropping from $1/2$ (SparseGPT) to $3\rho/2$ (Block-OBS-GS) where $\rho=1-s$, hence strictly lower for sparsity $s>2/3$; per-row reconstruction error at most SparseGPT's, strict whenever the activation Hessian has non-zero off-diagonal coupling; and a closed-form distance bound to the unregularised joint-OBS optimum that contracts geometrically per sweep. Across six Llama-3 and Qwen-3 backbones from 3B to 70B parameters, Block-OBS-GS attains the lowest WikiText-2 perplexity at every tested sparsity on the three Llama-3 backbones, with a $47\%$ perplexity reduction over SparseGPT on Llama-3.1-8B at $s=0.875$; the highest seven-task commonsense Avg7 on six of twelve populated cells; 5-shot Massive Multitask Language Understanding (MMLU) within $0.026$ of SparseGPT across populated cells; and the lowest WikiText-2 perplexity on four of five backbones under $2{:}4$ structured sparsity.

Increasing the replay ratio, where an agent's network is updated multiple times per environment interaction, is an effective strategy for improving sample efficiency in reinforcement learning. However, its effects on the network capacity of multiagent reinforcement learning (MARL) are not yet well investigated. In this paper, we show that high replay ratios induce a severe dormant neuron problem in the centralized global Q-network of MARL, where a large fraction of neurons become inactive, thereby reducing network capacity and destabilizing learning. To address this problem, we propose Ensemble Reset (EnSet) to stabilize MARL training at high replay ratios. First, guided by theoretical analysis, EnSet employs an ensemble of global Q-value networks with periodic resets to mitigate neuron dormancy during frequent updates. Second, EnSet diversifies replay experience by exploiting multiagent translation invariance as an inductive bias in the global Q-value function to prevent overfitting. Extensive experiments in SMAC, MPE, and SMACv2 environments demonstrate that EnSet consistently improves various MARL algorithms at high replay ratios with fewer environment steps.

Reinforcement learning with verifiable rewards (RLVR) has advanced the reasoning capabilities of large language models (LLMs). However, existing RLVR methods often suffer from exploration inefficiency due to mismatches between problem difficulty and model capability: overly difficult problems hinder reasoning path discovery, while overly simple problems offer little learning signal. To address this, we first formalize the effect of problem difficulty by quantifying the relationship between loss descent magnitude and rollout accuracy. Building on this analysis, we propose SEELE, a supervision-aided RLVR framework that dynamically adjusts problem difficulty to lie within the high-performance region. SEELE augments each training sample by appending a hint (part of a full solution) for difficulty reduction. Unlike previous hint-based approaches, SEELE deliberately computes the hint length for each individual problem to achieve an optimal difficulty. The optimal hint length is determined via multi-round rollout sampling, where an item response theory model fits accuracy–hint pairs from previous rounds to predict the next-round hint. This instance-level, real-time difficulty adjustment aligns problem difficulty with the evolving model capability, thereby improving exploration efficiency. Experiments show that SEELE outperforms Group Relative Policy Optimization (GRPO) and Supervised Fine-tuning (SFT) by +10.0 and +8.4 points, respectively, and exceeds the best prior supervision-aided approach by +3.8 points on average across six math reasoning benchmarks.

Low-label transfer is often decided by a practical question: after a representation looks useful, how many labels are still needed to learn the downstream boundary? Existing transferability and probe-quality scores rank global alignment, likelihood, margins, or calibration, but they do not measure how much target data remains close to the fitted decision boundary, where scarce labels are most costly. We propose effective boundary complexity, a label-budget diagnostic for frozen representations and linear probes. It measures mass in a thin strip around the learned boundary and normalizes it by class balance, so dense ambiguous regions and rare-class difficulty are both counted. We prove that, under a smooth representation map, this boundary mass does not change because of generic volume expansion: density change and tangential surface change cancel at first order, leaving only stretch or compression in the boundary-normal direction. This yields a cached-feature estimator that fits a probe on one split, counts held-out features near its boundary, and avoids high-dimensional density estimation. Experiments support the diagnostic at three levels. Controlled transformations match the predicted law across 74 maps with 0.38% median relative error. On real frozen features, normal-direction changes move both measured boundary mass and label demand, while tangent changes do not. A differentiable surrogate reduces measured complexity by 16–36% at matched top-1 utility. Finally, CIFAR-100 and Tiny-ImageNet audits show boundary mass acting as a second-stage diagnostic for near-tied representations, with external multiclass results showing how label-grid saturation limits policy-level gains on label grids where many tasks tie.

Parameter-efficient fine-tuning of quantized large language models is dominated by QLoRA, which combines 4-bit weight quantization with FP16 low-rank adapters. We show the FP16 adapter precision is unnecessary: a single binary bit per adapter parameter, combined with a learned per-layer scale $s$ and a Scaled Straight-Through Estimator, matches or exceeds QLoRA across a broad range of benchmarks while reducing adapter storage by $14$--$16\times$. We introduce two variants--NormScale, which clips STE gradients to the unit ball, and LearnThresh, which makes the binarization threshold learnable---unified under a framework we call BQ-LoRA. Our main theoretical contribution is a tight characterization of the BQ-LoRA hypothesis class: it has bounded entrywise norm, a high-probability spectral norm bound that is $\sqrt{r}$ times sharper than the worst case, a generically full-rank update, and consequently a Rademacher complexity that yields an explicit $\tilde{O}(s_{\max}\sqrt{d_{\text{in}} d_{\text{out}}/n})$ generalization bound. None of these properties hold a priori for FP16 LoRA. On Qwen2.5 (1.5B/7B/72B) and LLaMA-2 (7B/70B), BQ-LoRA matches or beats QLoRA and LoftQ at $\sim 1/16$ the adapter footprint; the predicted regularization signature appears empirically in dataset ablations and training-loss curves. We extend Punica's SGMV kernel with a binary inner-product primitive (BQ-SGMV) and prove a cache phase-transition result: shrinking adapters $16\times$ enlarges the HBM cache from $2{,}500$ to $39{,}000$ adapters, eliminating misses under skewed workloads.

Brain signals exhibit substantial multi-center heterogeneity, strong inter-subject variability, and highly diverse task settings, which makes it difficult to build models that generalize beyond a single task. While recent brain signal foundation models leverage large-scale pretraining to learn generalizable representations, they often fall short in practice because they still require target-data-specific supervised adaptation, lack multi-task zero-shot capability, and rarely provide verifiable evidence to support their predictions. In this work, we present BrainCoT, a multi-task zero-shot Brain signal foundation model with neurometric-anchored Chain-of-Thought (CoT) reasoning. BrainCoT unifies heterogeneous tasks into an instruction-driven framework so that, after pretraining, a single model can be directly applied to multiple tasks in zero-shot settings without additional adaptation. It further introduces neurometric-anchored chain-of-thought grounded in neurometric evidence for verifiable reasoning, and adopts Decision-Consistent Rationale Training to keep process learning aligned with correct predictions. Across extensive evaluations, BrainCoT achieves an average ACC of 67.41\% and AUC of 74.33\% in zero-shot classification without downstream training data, outperforming the strongest linear-probing baseline trained with 1\% labeled downstream data by 3.85\% in ACC and 5.97\% in AUC. Code will be released upon acceptance.

Naturalistic fMRI offers a time-resolved window into cortical dynamics during continuous multimodal experience, where future brain states are jointly governed by endogenous neural context and exogenous sensory drive. Forecasting such dynamics, however, remains challenging because fast-changing stimulus streams must be reconciled with the sluggish hemodynamic BOLD response, while cortical activity is organized across functionally heterogeneous brain networks. To address these challenges, we introduce BrainVista, a multimodal autoregressive framework for observed-stimulus conditional brain-state forecasting. BrainVista predicts future fMRI states from the history of brain activity and stimulus tokens temporally aligned to each prediction query, enabling autoregressive prediction without access to future ground-truth fMRI states. This design supports causal, temporally consistent modeling of naturalistic brain dynamics under realistic sensory stimulation. Specifically, we design Network-wise Tokenizers for functionally structured brain representations, a Spatial Mixer Head for cross-network interaction refinement, and Stimulus-to-Brain masking to preserve temporally causal stimulus–brain conditioning while preventing future brain-state leakage and stimulus look-ahead. We evaluate BrainVista on Algonauts 2025, CineBrain, and HAD, three large-scale naturalistic fMRI benchmarks, under long-horizon conditional autoregressive forecasting. BrainVista consistently outperforms baselines, improving pattern correlation by 10.8\% and 9.4\% relative to the strongest baseline on Algonauts 2025 and CineBrain, respectively.

Reinforcement learning with verifiers (RLVR) has become a central paradigm for improving LLM reasoning, yet popular group-based optimization algorithms like GRPO often suffer from exploration collapse, where the models prematurely converge on a narrow set of high-scoring patterns, lacking the ability to explore new solutions. Recent efforts attempt to alleviate this by adding entropy regularization or diversity bonus. However, these approaches do not change the \textit{winner-takes-all} nature, where rollouts still compete for individual advantage rather than cooperating for maximizing global diversity. In this work, we propose Group Cooperative Policy Optimization (GCPO), which shifts the training paradigm from rollout competition to team cooperation. Specifically, GCPO replaces independent rollout scoring with team-level credit assignment: a rollout is rewarded by how much it contributes to the team's valid solution coverage, rather than its individual accuracy. This coverage is described as a determinant volume over reward-weighted semantic embeddings, where only correct and non-redundant rollouts contribute to this volume. During advantage estimation, GCPO redistributes the collective team reward to each single rollout according to its average marginal contribution to the team. This cooperative training paradigm routes optimization toward non-redundant correct reasoning paths. Experiments across multiple reasoning benchmarks demonstrate that GCPO significantly improves both reasoning accuracy and solution diversity over existing approaches. Code will be made available.

Low-bit post-training quantization (PTQ) is a pivotal technique for deploying Vision-Language Models (VLMs) on resource-constrained devices. However, existing PTQ methods often degrade VLMs' accuracy due to the heterogeneous activation distributions of text and vision modalities during quantization. We find that this cross-modal heterogeneity is distributed unevenly across channels: a small subset of channels contains most modality-specific outliers, and these outliers typically reside in different channels for each modality. Motivated by this, we propose SplitQ, a channel-Splitting-driven post-training Quantization framework. At its core, SplitQ introduces a novel Modality-specific Outlier Channel Decoupling (MOCD) module that effectively isolates salient modality-specific outlier channels with minimal overhead. To further address the remaining cross-modal distribution discrepancies, we design an Adaptive Cross-Modal Calibration (ACC) module that employs dual lightweight learnable branches to dynamically mitigate modality-induced quantization errors.Extensive experiments on popular VLMs demonstrate that SplitQ significantly outperforms existing approaches across 6 popular multi-modal datasets under all evaluated quantization settings, including W4A8, W4A4, W3A3, and W3A2. Notably, SplitQ preserves 95.3% of FP16 performance under the challenging W3A3 setting (69.5 vs. 74.3), pushing the efficiency frontier for deploying advanced VLMs.

Group dance generation aims to synthesize coordinated multi-dancer choreography from music, with broad applications in animation and interactive content creation. This task requires modeling dense inter-person dependencies to ensure spatial coordination, while naturally preserving individual dancer identities. Existing approaches model all dancers jointly with end-to-end transformers, which tie the architecture to a fixed group size and entangle per-dancer identities across frames. We propose ChainDance, a scalable framework that reformulates group dance generation as a Chain-of-Dancers: a sequential decomposition over per-dancer conditional distributions, allowing a single model to scale across arbitrary group sizes without retraining and naturally preserving per-dancer identity. Built on a frozen single-dancer diffusion backbone, ChainDance introduces two lightweight modules: a Role-Aware Text Encoder (RATE) for per-dancer semantic conditioning, and a Group-Aware Motion Encoder (GAME) that aggregates previously generated dancers via a distance-weighted graph convolutional network, and incorporates a training-free noise optimization procedure at inference time to enforce global spatial coherence. Experiments on AIOZ-GDance demonstrate that ChainDance achieves state-of-the-art performance in motion quality and identity preservation, with 3-4 times fewer parameters and requiring 3-6 times less training time compared to prior approaches.

Distributionally robust optimization (DRO) provides a principled framework for learning under distribution shift, but its practical use is hindered by the difficulty of evaluating worst-case risks for nonconvex loss functions. We study a penalized DRO formulation in which the adversary may choose any distribution but incurs a Wasserstein penalty for deviating from the empirical distribution. We show that the adversary’s problem can be reformulated as an optimization problem over transport maps that push empirical samples to adversarial ones, and we prove that optimal maps are cyclically monotone. We also show that standard adversarial training---based on per-sample local optimization---violates cyclical monotonicity and wastes transport costs unless the adversary is severely restricted. We propose two remedies. First, we introduce multi-start particle ascent, which alternates parallel gradient ascent with reassignment to enforce cyclical monotonicity across samples. Second, we parameterize adversarial maps as gradients of input-convex neural networks, which guarantees cyclical monotonicity by construction. Experiments on robust regression, image classification, and robust control show that our methods consistently outperform standard adversarial training and state-of-the-art baselines, achieving improved robustness and better generalization under distribution shift.

Generative models are often required to serve many downstream conditional tasks. In the training-free setting, a pretrained unconditional model is reused as a prior and combined with a task-specific likelihood to sample from the corresponding posterior, without retraining the generative model. While such conditional samplers have been developed for diffusion and flow generative models, training-free conditional sampling for Schrödinger bridge (SB) generative models remains unexplored, despite the growing appeal of SBs as flexible stochastic transport models between general source and target distributions. We propose BridgeTwist, a training-free conditional sampler for pretrained SB generative models, formulated as a tempered twisted sequential Monte Carlo scheme. At each step, the sampler uses a closed-form plug-in twist surrogate, which combines the velocity and score fields exposed by SB pretraining into an explicit endpoint predictor. Since this surrogate is less reliable near the source, we use a time-varying temperature to temper the twist, downweighting its influence at early steps. Telescoping importance weights then cancel all intermediate twist factors, yielding an asymptotically exact conditional sampler without any auxiliary twist network or pilot rollouts. Empirically, BridgeTwist consistently outperforms training-free baselines on class-conditional sampling and inpainting on MNIST and CIFAR-10, and on text-to-image generation on CelebA-HQ.

Vision-language models such as CLIP enable zero-shot classification from arbitrary class names, but under test-time corruptions, their visual encoder projects distorted features onto incorrect textual anchors with high confidence—a phenomenon we term \textbf{modality bias}. Standard entropy minimization reinforces these overconfident errors, triggering a destructive confirmation-bias loop that collapses the predicted label distribution. We propose \textbf{CDIM} (Cross-Modal Information Maximization), an online test-time adaptation method that breaks this loop at three levels. At the \textit{representation level}, a frozen self-supervised DINO backbone is combined with CLIP via a cross-covariance logit ensemble; by projecting text prototypes into DINO's geometric space—rather than the reverse—the classifier operates free of text bias. At the \textit{optimization level}, a gated information-maximization objective (GT-IM) pairs per-sample sharpening with batch-level diversity, while an adaptive gate excludes unreliable samples from the gradient. At the \textit{prediction level}, an online logit adjustment corrects the residual class-distribution skew. Ultimately, \textbf{CDIM} validates its broad effectiveness by delivering robust, competitive performance across demanding setups, including continual TTA, mixed-domain shifts, and diverse domain generalization benchmarks.

Classical compute-optimal scaling laws assume an unbounded supply of fresh pretraining data, yet pretraining is increasingly entering the regime where compute is growing faster than high-quality data. We propose *Compute-Data (CD) scaling laws*, a unified framework that bridges *compute-optimal* scaling, in which data scales freely with compute, and *data-optimal* scaling, in which the corpus is fixed and compute can grow unbounded. CD scaling extends classic scaling by introducing a *token effectiveness function* that quantifies how much a *derived* token, produced for instance by multi-epoch repetition or paraphrasing, is worth relative to a fresh one, ranging from a perfect substitute to no value at all. Fitting $\eta$ for two data expansion strategies (multi-epoch repetition, paraphrasing) across model sizes ranging from 14M to 600M parameters on the Dolma-3 corpus, we find that it is far from constant: it depends jointly on model size, the tokens-per-parameter ratio, and the amount of derived data, and saturates as the corpus is expanded. The functional form of $\eta$ implies that substituting compute for data *diminishes* with both model size and data availability, and partitions training into three operational regimes: compute-bound, data-bound, and model-bound, showing that classic compute-optimal allocation is suboptimal across most of the practically relevant regime.

Reinforcement learning (RL) is essential for adaptive decision-making, e.g., robust Meta-RL and domain randomization (DR), and efficient post-training of large language models (LLMs), where tasks are commonly structured as Markov decision processes or prompts. A central bottleneck in all these settings is the cost of agent-environment interactions required for task selection and policy optimization. Model predictive task sampling (MPTS) has emerged as a promising approach to improve sample efficiency by querying informative tasks via a risk predictive model (RPM) trained on optimization history. However, existing RPMs suffer from two failure modes: (i) sparse optimization histories, which limit RPMs' generalization across the task space, and (ii) a one-step lag in risk approximation, which causes biased difficulty estimation as the policy evolves. This work formalizes these failure modes, derives lagged-risk decompositions that expose the resulting temporal bias, and recasts risk prediction as an in-context sequence modeling problem. The developed RPM integrates a one-step look-ahead mechanism with a temporal-weighted sliding window to mitigate data non-stationarity and scarcity, without requiring additional rollouts. Empirically, our approach yields more reliable difficulty estimates and consistent performance gains across robust Meta-RL, DR, and prompt curriculum RL post-training of LLMs.

Hybrid large language models (LLMs) integrating state space models (SSMs) and full attention face a critical caching bottleneck due to mismatched caching granularities. While full attention allows flexible, token-level cache reuse, SSMs are always restricted to coarse-grained block- or request-level reuse. Consequently, an SSM layer's cache miss can easily invalidates successful fine-grained hits in adjacent attention layers, disrupting the entire reuse pipeline. To resolve this, chunk-level position-independent cache (PIC) reuse is highly desirable to align the caching granularity across all layers. However, enabling PIC reuse for history-dependent SSMs remains fundamentally challenging. To bridge this gap, we propose HyPIC, the first unified framework achieving PIC reuse for hybrid architectures. We introduce residual state propagation and online decay correction, which efficiently adapt offline SSM chunk caches to new contexts by explicitly recomputing minimal boundary tokens. Furthermore, we design a fast parallel recomputation pipeline exploiting the linear superposition property of SSMs to eliminate sequential bottlenecks. Experiments on diverse hybrid models and datasets demonstrate that HyPIC achieves a 2$\times$ speedup in time-to-first-token (TTFT) while maintaining highly competitive accuracy.

Feed-forward 3D Gaussian Splatting has recently emerged as a promising alternative to per-scene optimization, enabling fast novel view synthesis by predicting Gaussian primitives directly from input images. However, it quickly faces a bottleneck of rendering efficiency while scaling from sparse to dense multi-view inputs, which offer more geometric constraints but produce a rapidly growing set of candidate Gaussians. Existing solutions cannot reliably distinguish redundant primitives, leading to many redundant primitives from repeatedly observed regions and increased rendering cost without much gain in visual quality. We propose BudSplat, a compact feed-forward 3DGS framework that learns to keep the most valuable Gaussians under a limited primitive count. BudSplat has two novel designs. The first is a detail-aware importance score that is extracted from multi-level visual tokens, providing an explicit cue for regions where pruning is likely to harm perceptual quality. The score can be injected into voxel fusion to preserve high-value local contributors during multi-view aggregation. The second is quality-aware Gaussian pruning that selects a compact set of primitives by jointly considering rendering contribution and detail importance. The quality-aware pruning enables local quota redistribution that helps avoid over-allocating primitives to large redundant regions. Experiments on RE10K, DL3DV, and ACID show that BudSplat achieves a favorable trade-off among rendering quality and efficiency.

Most coding-agent benchmarks ask whether generated code behaves correctly. That remains essential, but repository-level engineering is increasingly agent-managed: one agent writes a repository, and later agents inspect, audit, or extend it as working context. In that setting, a generated repository is not only an answer to a task but also a communication artifact for future work. Even when strong agents nearly satisfy the visible behavioral objective, repositories can differ in how clearly they expose the intended behavior and design choices behind that behavior. We introduce Build-and-Find, a protocol for evaluating whether downstream agents can recover those intended choices from generated repositories, and how much inspection that recovery requires. For each task, a builder sees a hidden repository specification and creates a codebase; a finder sees only the codebase and a specification-traced multiple-choice question bank. The protocol separates behavioral correctness from artifact-side recovery and reports recovery accuracy, repeatability, implementation coverage, and inspection effort. Accuracy and stability act as gates: effort is interpreted only when recovery succeeds reliably. Among artifacts from which the same intent can be recovered, lower effort by the same finder suggests that the artifact makes that intent easier to locate. Question-only and spec-only controls quantify generic priors and specification access, while audits separate omitted claims from finder failures and check whether correct answers cite artifact evidence. In the released high-prior task pack, recovery accuracy is near saturation, so inspection effort and finder-specific effects provide the main panel-local comparison. We release the harness, tasks, generated artifacts, records, canonical tables, reports, scripts, metadata, licenses, and evidence audit to support auditable extensions and private task packs.

Hyperbolic neural networks excel on hierarchical graph data, yet existing hyperbolic Transformers lift standard attention into curved space without using the ideal boundary, the structure most directly tied to tree hierarchy. We propose \textbf{Horospherical Self-Attention} (HSA): each query selects an ideal-boundary direction and scores keys by their Busemann depth along that direction, yielding a token-to-token hyperbolic attention kernel. We prove flat-limit consistency (HSA recovers dot-product attention as $K\to 0^-$), characterise score level sets as horospheres, and bound key-side gradients. HSA is integrated into \textbf{Busemannformer}, a full hyperbolic graph Transformer. Busemannformer-LD reaches \textbf{91.78\%} F1 on Disease-NC; in a controlled score ablation, its log-depth Busemann score achieves the best Disease-NC result among tested scores, reaching \textbf{93.29\%} F1 and performance comparable to the current SOTA Hypformer result with a smaller training budget. For link prediction, we find a score-symmetry principle: Busemannformer-LD with a symmetric score achieves \textbf{90.71\%} AUC on PubMed-LP ($+$10.3 pp over the asymmetric variant), while the asymmetric score performs best on Disease-LP among our variants and outperforms all Euclidean baselines. Ablations identify log-depth Busemann scoring as the key component and reveal a clear failure mode on Airport, where labels track centrality rather than tree depth.

The rapid growth of generative AI is reshaping the energy footprint of modern computing infrastructure. Hyperscale AI data centers now operate at power levels comparable to large industrial facilities, creating new challenges for grid reliability, energy procurement, and real time demand flexibility. At the same time, their large controllable loads, cooling systems, and on site storage make them potential participants in grid services such as frequency regulation. However, existing evaluation environments do not capture the coupled decision making required for grid interactive operation. We introduce C2G-BENCH, a cyber-physical benchmark for hierarchical reinforcement learning in hyperscale data centers. C2G-BENCH simulates a real-world scenario, where a 250 MW-class hyperscale data center participates in grid frequency regulation while serving realistic AI workload demand under weather-driven ambient conditions. The benchmark explicitly links 15-minute market level orchestration with 5-second physical control, enabling agents to jointly reason over regulation commitments, workload flexibility, cooling, battery dispatch, and facility safety. C2G-BENCH achieves this through two coupled Gymnasium environments: C2GMacroEnv, where a high level controller bids into the regulation market, and C2GFastEnv, where a low level controller actuates the Dynamic Voltage and Frequency Scaling (DVFS), Coolant Distribution Unit (CDU) pump speed, HVAC effort, and Battery dispatch (BESS). The simulator models a 250 MW class two zone facility using seven modular physics engines covering workload, thermal, electrical, battery, grid signal, weather, and market dynamics, driven by real workload and weather traces. We benchmark a family of rule-based, LLM-based & RL-based controllers for both the macro and fast environments, and evaluate them on four stress-test environments in multiple grid markets. Baseline results show that active low level control substantially improves tracking quality: a rule-based macro controller paired with a RL-based low level controller reduces the tracking RMSE from 2,737 kW to 229 kW compared with a macro only configuration, while maintaining full throughput and zero thermal violations. These results demonstrate that C2G-BENCH exposes meaningful cross layer tradeoffs and provides a reusable benchmark for studying safe, hierarchical control of grid interactive AI data centers.

Maintaining character consistency over long-horizon gameplay is a critical bottleneck for deploying Large Language Models (LLMs) as Non-Player Characters. Prevalent discrete Affective State Machines (ASMs) fail due to two paradoxes: (1) the \textit{Cliff-edge Effect}, where threshold-based switching causes jarring narrative discontinuities; and (2) \textit{Auto-regressive Inertia}, where accumulated dialogue history dilutes current state instructions. We propose \textbf{C3 (Causal-Continuous Characterization)}, which couples a continuous state space with \textit{hysteresis-aware dynamics} to dampen emotional volatility, and a \textit{State-Conditioned Memory Rewriting} mechanism that re-contextualizes past events through the lens of the current relationship. On a 100-turn dynamic simulation (\textsc{Chronos-Sim}, 5{,}000 turns per condition), C3 achieves a \textbf{10.5\% absolute SAR improvement} over a strong smoothed-ASM baseline (Welch's $t$-test, $p < 0.0001$, Cohen's $d \approx 3.97$). Cross-model evaluations on \texttt{Llama-3-8B-Instruct} and \texttt{Mistral-v0.2-7B} reproduce the relative gain, indicating the mechanism transfers across backbones although absolute SAR remains backbone-dependent. A human study with 20 game designers (ICC$=0.74$) corroborates higher perceived consistency, and a separate factual audit (Fleiss' $\kappa=0.81$) confirms the Rewriter alters interpretation without corrupting facts. Deployment analysis shows a $\sim$60\% reduction in long-session token consumption.

Despite the increasing scale of genome language models (gLMs), their ability to decode the function of regulatory sequences remains unclear. gLM pretraining relies on sequence reconstruction, which may struggle due to the noisy, rapidly evolving nature of regulatory DNA. Self-supervised contrastive approaches provide a promising alternative. Inspired by language-image architectures like CLIP, we demonstrate contrastive promoter-protein pretraining (C3P). By learning to align promoters to their corresponding proteins, we leverage the rich representations of proteins learned by protein language models as supervisory signal for the learning of promoter representations. After training on 88 million bacterial promoter-protein pairs, we evaluate the predictive power of C3P-learned promoter representations for inference of curated regulatory annotations, finding multi-fold improvement over leading gLMs. We also introduce zero-shot co-regulated gene retrieval, the ability to find co-regulated genes in a genome using no experimental data. We find that compared to a randomly initialized baseline, C3P training consistently provides significant zero-shot performance gains, unlike gLMs. Scaling analysis reveals the potential for further improvement as well as the efficiency of C3P, which achieved strong performance at a fraction of the training cost of leading gLMs. In addition to demonstrating that C3P training is effective for learning representations of bacterial regulatory sequences, our strong zero-shot co-regulated gene retrieval performance suggests the possibility of decoding gene regulation for millions of bacteria from their genomes alone.

Multi-agent LLM systems (MAS) improve reasoning by decomposing across specialized roles, but their text communication is expensive. Each agent decodes thousands of intermediate tokens for the next, and successive prefills grow cumulatively. Recent work shows that latent-space signals can substitute for explicit text between agents, cutting cumulative prefill cost. When the latent reasoning also replaces an agent's text reasoning, it additionally cuts intermediate tokens. Here, we find that the prefill cache alone suffices as the latent-space signal between agents, and the per-agent forward chains can be compressed into a single pass. With \emph{exactly one prefill and one decode per problem}, we sharply cut per-rollout tokens and prefill wall time. We call this architecture \textbf{CacheMAS}, in which prefill agents pass information through cache communication, and a final agent decodes the answer. As a bonus, \textbf{CacheMAS} is one autoregressive rollout, so unlike prior MAS, standard sequence-level RL can jointly optimize all agents at once without per-agent reward design. On math, multi-hop QA, and code, CacheMAS uses $\sim$$3\times$ fewer tokens, runs $\sim$$2\times$ faster, and after joint optimization beats prior MAS by $+1$ to $+8$ points on every benchmark. Cache communication is not just a token-efficient substitute for text, but a practical interface for joint MAS optimization.

Although recent training-free acceleration methods have achieved promising speedups for Diffusion Transformers, most rely on historical features, either by directly reusing cached features or by forecasting future features from them. Such a history-to-future design can accumulate errors when the prediction horizon becomes long. To address this issue, we propose FuCa, a Cache-the-Future training-free acceleration framework. Instead of extrapolating future information solely from the past, FuCa caches sparsely computed future features and uses them as online revision signals. FuCa revises the latent-state update and the corresponding velocity, and then recovers skipped timesteps through lightweight interpolation. It requires no extra training or model-specific auxiliary modules, while keeping the number of DiT forward evaluations low. Experiments on FLUX.1-dev, Qwen-Image, Wan2.1-1.3B, and Hunyuan Video show that FuCa achieves up to about 5$\times$ speedup for image generation and more than 4.3$\times$ speedup for video generation, with higher reconstruction quality than existing training-free baselines.

Evaluating generative models requires measuring not only sample fidelity but also how well a generated set covers the real data distribution. Recent work has therefore introduced recall-oriented metrics for distribution-level evaluation. However, existing recall metrics often reduce distributional coverage to geometry-based coverage. This simplifying assumption weakens their diagnostic value and interpretability in two ways: these metrics provide only a single recall score that cannot distinguish essential diversity failures such as mode dropping and mode collapse, and geometry-based coverage can provide unreliable estimates by assigning high recall to poorly covered generative distributions. We therefore introduce CADMA, a capacity-aware framework for recall-oriented evaluation and diagnosis. CADMA evaluates distributional coverage by formulating local real--fake allocation as a one-to-one matching problem, separating geometry-based coverage from effective coverage. This yields multi-dimensional recall diagnostics that quantify mode dropping and mode collapse while reducing the overly optimistic estimates produced by geometry-based recall metrics. Through extensive investigations on toy experiments and generative models, we show that CADMA provides more reliable recall estimates and more fine-grained, interpretable diagnostic signals of diversity failures than existing metrics.

Existing 3D scene-grounded Large Language Models (3D-LLMs) focus on answering questions grounded in simplified single-room 3D scenes, lacking the ability to reason over real-world household environments containing multiple interconnected rooms and diverse object categories. We introduce CAIRN, a topology-aware 3D-LLM for multi-room 3D scene understanding. CAIRN aligns transformer attention with scene hierarchy, giving the model explicit awareness of object-level relations and room-level connectivity. It enriches object tokens with room-local relational context via a graph neural network, introduces learned room tokens for room-level abstraction, and applies a hierarchical attention mask with geometric bias to route information according to scene topology. CAIRN is developed on CAIRN-MR, a benchmark we introduce on HM3D for multi-room 3D scene understanding, covering grounding, captioning, and four question-answering tasks that progressively evaluate from intra-room perception to cross-room reasoning. Experiments show that CAIRN outperforms prior 3D-LLMs by a large margin across all CAIRN-MR tasks while remaining competitive on five single-room benchmarks.

Generative models can produce individually plausible samples while deviating substantially from a reference set in the distribution of key features. For example, a model pretrained on broad drug-like chemical space may generate molecules whose molecular features differ from those of a therapeutic class of interest, such as known antibiotics. Correcting such distributional miscalibration is challenging: direct finetuning on the reference set can overfit and does not control which features are matched. To fill this gap, we introduce kernel Calibrating Generative Models (kCGM). kCGM minimizes a maximum mean discrepancy (MMD) between generated and reference feature distributions using an unbiased score-function estimator, with KL regularization to remain close to the pretrained model. On a reference set of 174 antibiotics, direct finetuning sacrifices chemical validity for feature-distribution matching, whereas kCGM improves target feature matching while increasing validity. We further demonstrate kCGM in protein and DNA generation tasks, showing it can adapt autoregressive, continuous-space diffusion, and discrete diffusion models using only feature-level supervision.

Memory facilitates question answering over long videos by extracting and retrieving facts to fit within the limited context windows of multimodal LLMs (MLLMs). Existing solutions typically extract independent memory entries from fixed-length video clips and thus cannot capture high-level semantics that need to be summarized over extended time periods, such as character traits and relations. Moreover, they rely solely on similarity-based retrieval and may fail to retrieve the fine-grained details required for question answering. To tackle these problems, we propose CAM, featuring continuous extraction for high-level semantics and \textit{adaptive querying} for fine-grained details. In particular, CAM stores the entities and relations extracted from video clips in a knowledge graph. To capture the high-level semantics of each entity or relation, CAM summarizes the local subgraph of the target entity or relation once the subgraph reaches a predefined size. To retrieve the fine-grained details required for question answering, CAM supports multiple search methods, such as knowledge graph traversal, video re-watching, and audio listening, and utilizes a planner-executor-verifier pipeline to adaptively compose these search methods according to question intent. Evaluations on three benchmarks show that CAM outperforms SOTA baselines and improves their accuracy by up to 23 percentage points.

Multimodal Recommender Systems (MMRS) leverage rich item features to mitigate data sparsity, yet their success largely hinges on the foundational homophily assumption. This paradigm makes existing models adept at recommending similar substitutes but poorly equipped to handle heterophilic complementarity. Consequently, complementary items often become geometrically isolated, leading to fragmented “Similarity Islands” and semantic drift when standard propagation is applied. To overcome these limitations, we propose MAGE (Manifold-Augmented Graph Embedding), a novel framework that decouples complementarity modeling into offline geometric warping and online topological propagation. The offline phase enhances item representations by extracting complement-aware residual directions via a Soft Conceptor mechanism and injecting them under a local gap-preserving constraint, enabling the model to capture functional relationships while reducing the risk of semantic drift. Synergistically, the online phase employs an asymmetric dual-stream GNN that treats visual features as stable anchors and augmented textual features as flexible probes, fused through a self-calibrated cross-modal mechanism to reconcile shared semantics and modality-specific residual cues. Extensive experiments on four Amazon datasets show that MAGE yields consistent improvements over strong multimodal recommendation baselines. Together with the supporting analyses in the appendix, these results suggest that geometry-aware augmentation can alleviate the semantic mismatch that arises in complementary recommendation settings. Our code is available at https://anonymous.4open.science/r/mage-4C60.

Large language models (LLMs) are increasingly used in high-stakes settings where they are expected to justify their predictions, and a common approach to understanding LLM predictions is via self-generated free-text explanations. However, free-text explanations are instance-wise and non-executable, making it hard to evaluate their alignment with model behavior or explain dataset-level behavior in a compact, human-interpretability manner. We study code as an executable alternative to free-text, enabling automatic evaluation and dataset-level explanation, and ask whether LLMs can \emph{consistently} explain their behaviors with code. This automatic verification enables a generation procedure that samples multiple candidate programs and selects the most consistent one. Using three open-source code-capable models and eight tasks with varying levels of linguistic difficulty, we study when LLMs produce consistent code explanations. We find that models can produce highly consistent code explanations in a subset of settings, but strong consistency is not the norm. Consistency is substantially lower for tasks involving more complex linguistic components, where performance is often near random guessing across models. In these harder settings, generated programs are longer but not more structurally complex, and increasingly rely on hard-coded linguistic patterns from the input. These findings suggest two possibilities: either LLMs are not good at producing compact explanations of their behavior, or some behaviors require dense, distributed computation that cannot be explained by compact, human-interpretable programs. Finally, we find that model-internal representations do not reliably predict when code explanations will be consistent, suggesting that LLMs do not know when they can or cannot explain their own behavior with code. Our work highlights code as a compact and verifiable alternative to free-text explanations, while also revealing limitations in using compact, human-understandable explanations to fully understand LLM behavior.

Text-to-SQL systems are typically evaluated by query-level execution correctness, but this terminal signal provides little guidance about which intermediate SQL decision caused success or failure. Token-level dense supervision is also ill-suited: SQL tokens do not align with complete semantic decisions, can penalize execution-equivalent queries, and are difficult to label reliably at scale. We therefore propose CAPER, which automatically derives clause-level supervision via counterfactual intervention on the SQL abstract syntax tree (AST), enabling root-cause error localization for reward modeling; the resulting data is used to train CAPER-9B, a lightweight Clause-PRM that provides clause-boundary feedback for policy optimization and candidate verification. Experiments on BIRD and Spider show that clause-aligned supervision not only improves execution accuracy, achieving up to a 15.3\% relative EX improvement over GPT-5.4, but also strengthens failure-localization capability, reaching 84.53\% accuracy and 90.60\% MRR on held-out failures.

The rapid advancement of Large Language Models (LLMs) based on Mixture-of-Experts (MoE) architecture has enhanced a growing demand for secure inference frameworks that protect both client inputs and server model weights. However, existing secure MoE inference frameworks suffer from two limitations: (1) a costly two-stage secure routing pipeline that uses expensive secure Top-$K$ and secure equality test protocols, resulting in significant computational and communication overhead, and (2) secure evaluation of nonlinear activations like SiLU requires several multiplication and comparison operations, resulting in significant communication overhead. In this paper, we present Capricorn, a highly efficient and secure MoE inference framework that overcomes the two limitations above. Firstly, Capricorn utilizes a novel one-stage routing pipeline that reduces both the total number of Top-$K$ protocol invocations and their input data size. At the same time, we propose a lightweight method to generate one-hot token selection matrices locally, thus avoiding expensive secure equality tests. Secondly, we design a secure and precise SiLU protocol, which uses secure lookup tables (LUTs) with dynamic bit-width and precision to reduce communication overhead. Extensive experiments demonstrate that Capricorn achieves up to $50.93 \times$ and $2.10 \times$ speedup in secure routing and secure SiLU protocols, respectively, compared to the state-of-the-art (SOTA) framework CryptoMoE (NeurIPS'25), while maintaining the original model inference accuracy.

Reinforcement learning with verifiable rewards (RLVR) has become the dominant paradigm for post-training large language models (LLMs) on complex reasoning and multi-step agentic tasks. However, the standard scalar reward signal is sparse and uninformative: it conveys whether a trajectory succeeded but not which reasoning steps failed or how to correct them, severely limiting sample efficiency on long-horizon tasks where successful trajectories are rare. We propose \textbf{CARD} (\textbf{C}ritique-\textbf{A}ugmented \textbf{R}einforcement \textbf{D}istillation), a training framework that amplifies the learning signal from a \texttt{verify_answer} tool within the GRPO objective via two complementary signals. First, a token-level self-distillation reweighting measures how much each token's generation was shaped by the expert critique: tokens that merely echo the critique text are masked, while critique-driven revision tokens are amplified. Second, a rejection-sampling distillation loss maximizes the likelihood of critique-free trajectories for responses where the critique provably improved the answer, directly accelerating the internalization of critique-guided behavior. Together, these two signals distill critique-augmented reasoning into the model's base policy without requiring additional rollouts or separate reward models. Experiments on four long-form deep search benchmarks and three multi-hop question-answering benchmarks demonstrate consistent improvements over the GRPO baseline. On the deep search benchmarks, CARD improves average performance by 40\% over the Qwen3-8B base model and by 6\% over GRPO; with Qwen3-14B, the gains are 46\% and 7\%, respectively. These results establish expert critique as an effective and scalable source of rich text supervision for reinforcement learning of long-horizon research agents. Furthermore, we show that \texttt{verify_answer} can serve as a test-time self-revision scaffold: at inference, the tool returns only task rubric criteria---without expert feedback---prompting the model to reflect on its draft and self-revise, yielding consistent performance gains over the no-VA baseline.

Multimodal Large Language Models (MLLMs) have demonstrated strong performance on vision-language tasks, yet their internal visual representations remain difficult to interpret. Sparse Autoencoders (SAEs) provide a scalable way to decompose dense model activations into sparse, interpretable features. However, existing SAE architectures primarily recover flat feature dictionaries and are less suited for explicit multi-level concept organization. In this paper, we introduce cascaded sparse autoencoders (CSAEs) for learning hierarchical visual concepts in MLLMs. Rather than nesting or stacking SAE sparse activation codes, CSAEs train a second-level SAE directly on the decoder weights of the first-level SAE, treating learned low-level feature directions as inputs for higher-level abstraction. This design enables CSAEs to learn “concepts of concepts” while avoiding drawbacks from the shared-prefix coupling of nesting, Matryoshka-style hierarchies and the bottlenecks of naively stacked SAEs. Experiments across Qwen3-VL, Gemma-3, and LLaVA on multiple visual datasets show that CSAEs improve interpretability in terms of hierarchical concept coherence over state-of-the-art SAE baselines. Results on concept steering further demonstrate that the learned concept groups support effective group-level interventions in MLLM outputs.

Quantifying variability in a target population relative to a reference population is central to many scientific and clinical problems (e.g., diseased vs. healthy). Yet, without paired data and in the presence of heterogeneous target variation, existing methods struggle to separate multiple modes of target-specific variation. We propose CASL-VAE, a deep contrastive latent variable model that learns structured latent generative factors from unpaired data. CASL-VAE factorizes variation into continuous common latent factors shared across populations and hierarchical salient latent factors that model target-specific heterogeneity as discrete subtypes and continuous within-subtype variation. Using variational inference, we show how approximate joint likelihood optimization over reference and target domains can be performed using unpaired data, providing a principled basis for paired-sample generation and cross-domain analysis. We validate CASL-VAE on semi-synthetic neuroimaging data, demonstrating improved subtype recovery and paired-sample generation compared to baseline clustering and generative models. We also validate its ability to reveal biologically plausible heterogeneity in Alzheimer's disease. Code will be released upon acceptance.

Autoregressive video generation aims at real-time, open-ended synthesis. Yet, cinematic storytelling is not merely the endless extension of a single scene; it requires progressing through evolving events, viewpoint shifts, and discrete shot boundaries. Existing autoregressive models often struggle in this setting. Trained primarily for short-horizon continuation, they treat long sequences as extended single shots, inevitably suffering from motion stagnation and semantic drift during long rollouts. To bridge this gap, we introduce \textbf{CausalCine}, an interactive autoregressive framework that transforms multi-shot video generation into an online directing process. CausalCine generates causally across shot changes, accepts dynamic prompts on the fly, and reuses context without regenerating previous shots. To achieve this, we first train a causal base model on native multi-shot sequences to learn complex shot transitions prior to acceleration. We then propose Content-Aware Memory Routing (CAMR), which dynamically retrieves historical KV entries according to attention-based relevance scores rather than temporal proximity, preserving cross-shot coherence under bounded active memory. Finally, we distill the causal base model into a few-step generator for real-time interactive generation. Extensive experiments demonstrate that CausalCine significantly outperforms autoregressive baselines and approaches the capability of bidirectional models while unlocking the streaming interactivity of causal generation.

Vision-Language Models (VLMs) have achieved strong performance on general multimodal reasoning, yet remain challenged in integrating nonlocal visual information to support semantically underdetermined visual reasoning. We describe this challenge as Fragmented Visual Reasoning. To this end, we propose Credit Assignment for Visual Evidence (CAVE), a structured process-reward method based on GRPO for interleaved visual reasoning. Specifically, CAVE evaluates the contribution of intermediate steps at the action level via three complementary reasoning process signals: belief update, evidence acquisition, and adaptive focus control, thereby guiding the model to optimize each reasoning action and learn more reliable visual reasoning strategies. Meanwhile, we construct TRACER-Bench, which covers four nonlocal and semantically confusable reasoning dimensions and provides key intermediate evidence to supervise reasoning paths. Experiments demonstrate that CAVE substantially improves performance on tasks requiring fragmented visual evidence integration, covering both public benchmarks and our newly introduced TRACER-Bench, while retaining competitive performance on general multimodal evaluations. Further analyses reveal that CAVE effectively improves the visual reasoning capacity and exhibits stronger robustness under longer-range and deeper cross-region dependencies. Our code and the proposed benchmark are available at: https://anonymous.4open.science/r/Anonymous_code-2D8B/.

Can we trust evaluation scores to capture an LLM's true real-world performance? *Certifiable* evaluation answers this question by providing guarantee for LLM evaluation. In particular, existing methods sequentially curate evaluation samples and keep updating confidence intervals (CIs) that cover the true performance with high probability (e.g., 95%) until some conditions are satisfied, e.g., the CI width reaches a target precision. However, existing methods are not generally anytime-valid: the claimed coverage (e.g., 95%) may fail when CIs are repeatedly updated and used to decide when to stop, leaving a *gap* between theoretical rigor and practice. This paper bridges this gap by proposing **Celeus**, a **C**ertifiable framework for **E**fficient **L**LM evaluation, which leverages ***E***-processes to build anytime-valid CIs. Concretely, we propose signals that combine two ingredients: (i) **U**ncertainty-guided sampling to select informative samples for evaluation, and (ii) **S**urrogate-assisted approximations for unevaluated samples. We prove that such signals remain unbiased for the evaluation score conditional on the past, enabling statistically grounded and anytime-valid $e$-process CIs. More importantly, the two ingredients reduce estimation variance and help reach the target precision with fewer evaluated samples. We also prove that CIs obtained by **Celeus** can shrink at a near-parametric rate up to logarithmic factors and analyze the oracle variance-optimal sampling rule that motivates the empirical uncertainty-guided one. Experiments show that **Celeus** reaches the target precision using 54-62% fewer evaluated samples than baselines, while preserving anytime-valid coverage.

This paper studies the problem of uncertainty quantification in large language models (LLMs), which is critical for reliable LLM deployment under the risk of hallucination. Existing approaches typically treat the response sets as unordered collections and rely on local pairwise similarity or clustering, which overlook relational signals among responses as well as their corresponding high-order structural semantics. Towards this end, we propose a novel approach named Complementary Signed Graph Propagation (CHAIN) for uncertainty quantification in LLMs. The core idea of CHAIN is to capture complementary signals among responses using a signed graph and then extract high-order structural information using random propagation. In particular, CHAIN builds a semantic graph where each response is considered as a node and edges represent complementary views, namely enrichment relations and contradiction relations. To acquire high-order structural signals, we update hidden states through graph-based information propagation along edges and channel fusion guided by randomized matrices. After propagation, these node states are summarized into graph-level scores for reliable uncertainty quantification. Extensive experiments on five datasets validate the effectiveness of the proposed CHAIN in comparison with competing baselines.

The attention mechanism in Transformers links queries, keys, and values through a sequence of computations, while the principles that constrain its form remain unclear. This paper examines attention under natural requirements that formalize basic properties of attention computation as transformations between representation spaces, and asks which forms of attention computation are admissible. We show that similarity computation, score-to-weight mapping, and value aggregation are unified by dual structures induced by convex potential functions and their Legendre transforms. In particular, the standard inner-product score is recovered as a limiting case of a potential-based score, the score-to-weight map must take a gradient-map form under structural requirements on attention normalization, and the value-aggregation rule must take a dual-coordinate barycentric form under structural requirements on representative-point aggregation. These results show that the standard components of attention are structurally constrained by the underlying dual structure, while the choice of potentials provides explicit degrees of freedom for understanding and modifying attention mechanisms.

Chains‑of‑thought (CoT) allow language models to verbalise multi‑step rationales before producing their final answer. While this technique often boosts task performance and offers an impression of transparency into the model's reasoning, this position paper argues that rationales generated by current CoT techniques can be misleading and are neither necessary nor sufficient for trustworthy interpretability. We analyze faithfulness as whether CoTs are not only human-interpretable, but also reflect the model’s underlying reasoning well enough to support responsible use. By synthesizing evidence from prior work, we show that verbalised chains are frequently unfaithful, diverging from the true hidden computations that drive a model's predictions. As CoT is increasingly relied upon in application domains, including high-stakes ones such as medicine, law, and autonomous systems, we further argue that this unfaithfulness is not taken seriously enough in these settings: our analysis of 1,000 recent CoT-centric papers finds that approximately 25\% explicitly treat CoT as an interpretability technique, including papers in high-stakes domains which heavily hinge on such interpretability claims. Building on prior work, we make three proposals: (i) avoid treating CoT as being sufficient for interpretability without additional verification, while continuing to use CoT for its communicative benefits, (ii) adopt rigorous methods that assess faithfulness for downstream decision-making, and (iii) develop causal validation methods to ground explanations in model internals.

Generative models excel at synthesizing high-fidelity samples from complex data distributions, but they often violate hard constraints arising from physical laws or task specifications. A common remedy is to project intermediate samples onto the feasible set at each sampling step. However, repeated projection can disrupt the learned sampling dynamics because this feasible set is defined for ''clean'' samples rather than noisy intermediate states. Thus, recent studies first estimate ''clean'' samples and then project them to the feasible set, but this increases algorithmic complexity and accumulates errors across steps. To address these issues, this paper proposes Chance-constrained Flow Matching (CCFM), a training-free method that formulates constraint enforcement during sampling as stochastic optimization while retaining the hard-constraint feasibility guarantee of projection-based methods. By avoiding direct projection of noisy intermediate states onto the clean-sample feasible set, CCFM mitigates the distributional distortion. Unlike approaches that first estimate and project ''clean'' samples, CCFM avoids complex multi-stage procedures. Experiments show that CCFM outperforms current state-of-the-art constrained generative models in modeling complex physical systems governed by partial differential equations, molecular docking problems, and physics-informed motion generation, delivering superior feasibility and fidelity.

Channel-adaptive Vision Transformers (ViTs) struggle to scale to non-RGB images with many uncorrelated or weakly correlated channels. Existing approaches roll out each channel into its own patch embedding, so the token sequence length grows linearly with the number of channels and self-attention cost grows quadratically. We introduce \emph{ChannelMixer}, a multi-channel tokenizer that aggregates all channels within each spatial patch into a single channel-mixing token, decoupling the token sequence length from the number of channels. ChannelMixer is pretrained once with a lightweight autoencoder in a channel-agnostic manner and the resulting tokenizer can be easily integrated with ViTs trained under both supervised and self-supervised objectives. On multi-channel imaging benchmarks spanning 3 to 29 channels, ChannelMixer matches or outperforms channel-adaptive state-of-the-art methods while using approximately $20\times$ fewer FLOPs and $29\times$ fewer tokens than rolled-out tokenizers, measured on 29-channel images with a ViT-S/16.

LLM agents now negotiate, allocate resources, and make commitments on a user's behalf. The vulnerability is not only in the final move but in what comes before it: in pre-action conversation, a model can be primed to make promises, lock in commitments, or reveal intentions that a counterpart will exploit when the action is finally taken. Yet standard game-based evaluations usually observe only the final move, leaving the speech-action gap unmeasured. We study this gap directly by introducing HL-PACT, a controlled platform with matched human-human, human-LLM, and LLM-LLM sessions across five repeated games: Prisoner's Dilemma, Battle of the Sexes, Ultimatum, Kuhn Poker, and Liar's Dice. Across 3,752 sessions, we randomize free-form pre-round chat and opponent identity masking, classify each message by its strategic content, and compare what players say, how they say it, and when the two diverge. We find that LLMs consistently activate the communication channel: mixed pairs contain more proposals, promises, and strategic reasoning than human pairs, with humans adapting to the LLM's negotiation register. This improves equilibrium selection in coordination games and shifts bargaining toward fairer offers. In games with profitable unilateral deviation, however, the same communication structure can create asymmetries: when identity is known, human partners selectively adapt to LLMs’ willingness to follow through on stated intentions, producing measurable exploitation. Masking reduces this asymmetry and makes the communication benefit more symmetric. The strategic value of language therefore depends on game structure, counterpart identity, and whether speech is separated from action. Evaluating LLM agents without communication can miss a central deployment risk: models may not fail because they reason poorly, but because their reliability is predictable.

Post-discharge risk prediction from electronic health records (EHRs) is difficult because many dependencies that link discharge-time observations to downstream complications, such as comorbidity cascades and drug-disease interactions, are absent from the record. External medical knowledge graphs (KGs) can supply these missing dependencies, but tracing them demands three properties: KG exploration must remain cost-bounded, retrieved evidence must be differentiated by source quality, and the resulting rationale must be citable for retrospective review. Large language models (LLMs) can plan and verify over structured evidence, making them natural candidates for KG reasoning, but existing LLM-based methods do not satisfy these three properties jointly. In this paper, we propose BAR, a Budget-Aware LLM Reasoning framework over medical KGs with three components. First, BAR refines the raw KG into disease-specific evidence graphs whose edges carry support scores and provenance records, turning the KG into a quality-annotated reasoning space rather than a static feature source. Second, an LLM then reasons over this graph through a plan-navigate-verify loop that decomposes the question into steps, retrieves evidence under a patient-specific budget, and revises when verification fails. Third, a reasoning policy is trained with a reward that compares predictions with and without acquired evidence, combined with acquisition cost and citation-integrity terms. Across 8 diseases and 3 prediction horizons on MIMIC-III and MIMIC-IV, BAR improves AUPRC by 3.4 points over the strongest baseline, raises citation precision from 59.8\% to 77.9\%, and consumes only 62-65\% of the budget cap. The code is available at https://anonymous.4open.science/r/BAR/.

Clinical EEG interpretation requires reasoning over full EEG sessions and integrating signal patterns with clinical context. Existing EEG foundation models are largely designed for short-window decoding and do not incorporate clinical context. We introduce CLEF, a clinically grounded long-context EEG foundation model. CLEF represents EEG sessions as 3D multitaper spectrogram tokens, enabling tractable Transformer modeling at session scale, and aligns embeddings with neurologist reports and structured EHR data through contrastive objectives. We evaluate CLEF on a new 234-task benchmark spanning disease phenotypes, medication exposures, and EEG findings, with more than 260k EEG sessions from over 108k patients. CLEF outperforms prior EEG foundation models on 229 of 234 tasks, improving mean AUROC from 0.65 to 0.74. Reconstruction-only pretraining surpasses prior EEG foundation models, while report and EHR alignment yields further gains. Held-out concept and external-cohort experiments suggest that these representations transfer beyond observed alignment targets. These results support session-scale, clinically grounded representation learning as a promising foundation-model paradigm for clinical EEG.

Despite impressive progress, existing feed-forward stereo 3D reconstruction models produce a single, irrevocable prediction. When these models fail, there is no mechanism to correct their mistakes. We introduce Click3R, a framework for interactive geometric correction of feed-forward models using sparse human correspondence clicks. Given an image pair and a small set of cross-view point correspondences provided by a user, Click3R injects geometric constraints into the 3D reconstruction networks through a lightweight point correspondence adapter, enabling the model to resolve ambiguities in challenging scenarios such as repetitive textures, symmetric structures, large viewpoint changes, or textureless regions. To evaluate interactive reconstruction, we curate a challenging benchmark specifically designed to expose failure modes of existing feed-forward stereo 3D reconstruction methods. Experiments show that Click3R dramatically reduces reconstruction error with as few as a single correspondence click, improving performance on both standard benchmarks and our curated dataset. Our results demonstrate that sparse human interaction provides an effective and practical mechanism for correcting geometric reconstruction errors.

Open-source web agents are now strong enough to execute realistic browser tasks, but training them with reinforcement learning still depends on weak supervision: binary task success is too sparse for credit assignment, while frontier-language-model judges are too expensive to call at every step and cannot be assumed available at deployment. We introduce CLIFT, a training and test-time scaling method built around conformal self-verification. During training, the agent answers natural-language verification questions about its own rollouts; a Compositional Conformal Certifier keeps only question signals whose URL-conditional evidence agrees with a training-time judge, assigns signed trust weights through polarity-aware lift, and blends the resulting verifier score into per-step rewards in a way that never subtracts from the judge baseline. At test time, the same certified bank is frozen and reused as structured evidence for Conformal Trajectory Selection (CTS): the agent samples a greedy rollout and one or more diverse retries, the self-verifier summarises each URL trace, and a conservative majority-vote rule chooses whether to swap away from the current incumbent without calling any external judge. This single mechanism supports three settings. On WebArena Infinity, CLIFT achieves state-of-the-art performance among open-source web agents. On VisualWebArena, a bank trained with the open model transfers to GPT-5.5 at test time and reaches state-of-the-art performance under the canonical harness. On Online Mind2Web, without training an agent on the benchmark, translating the certified question bank improves a live-web agent in zero-shot evaluation. Together these results position conformal self-verification as a way to turn costly judge feedback into a reusable training signal and a judge-free test-time scaling signal.

Clinical diagnosis begins with doctor-patient interaction, during which physicians iteratively gather information, order examinations, and refine diagnosis through patients’ responses. This dynamic reasoning process is poorly represented by existing LLM benchmarks that focus on static question-answering. To mitigate these gaps, recent methods explore dynamic medical frameworks involving interactive clinical dialogues. Although effective, they often rely on limited, contamination-prone datasets and lack granular evaluation. In this work, we propose ClinMAS, a knowledge-grounded multi-agent simulation framework for evaluating clinical reasoning in LLMs. Grounded in a disease knowledge graph, our method dynamically generates patient cases and facilitates multi-turn interactions between an LLM-based doctor, an automated patient agent, and an examiner agent. Our evaluation goes beyond diagnostic accuracy by incorporating fine-grained efficiency analysis and rubric-based assessment of diagnostic quality. Experiments show that ClinMAS effectively exposes critical clinical reasoning gaps in state-of-the-art LLMs, offering a more nuanced and clinically meaningful evaluation paradigm.

The widespread adoption of AI assistants has prompted the development of privacy-aware platforms designed to extract insights from real-world usage. Their privacy protections primarily rely on layering multiple heuristic techniques, such as PII redaction, clustering, and aggregation. In this paper, we put their privacy claims to the test by presenting CLIOPATRA, the first attack against privacy-preserving LLM-based insights systems. Our attack involves an adversary that carefully designs and inserts malicious chats into the system to break multiple layers of protections and induce the leakage of sensitive information from a target user’s chat. We evaluate CLIOPATRA on one such platform, Anthropic’s Clio, and target synthetically generated medical chats to show that an adversary can successfully and confidently (with nearly 100% precision) extract the medical history contained in these chats in up to 65% of cases. We also show that CLIOPATRA can stealthily extract information by obfuscating the private information in the generated insights. Finally, we demonstrate that existing ad hoc mitigations, such as LLM-based privacy auditing, are unreliable and fail to detect major leaks. Taken together, our findings indicate that, even when layered, current heuristic protections are insufficient to adequately protect user data, and that prompt injection has been an understudied risk in LLM-based insight systems.

Training implicit neural representations (INRs) to capture fine-scale details typically relies on iterative backpropagation and is often hindered by spectral bias when the target exhibits highly non-uniform frequency content. We propose ELM-INR, a backpropagation-free INR that decomposes the domain into overlapping subdomains and fits each local problem using an Extreme Learning Machine (ELM) in closed form, replacing iterative optimization with stable linear least-squares solutions. This design yields fast and numerically robust reconstruction by combining local predictors through a partition of unity. To understand where approximation becomes difficult under fixed local capacity, we analyze the method from a spectral Barron norm perspective, which reveals that global reconstruction error is dominated by regions with high spectral complexity. Building on this insight, we introduce BEAM, an adaptive mesh refinement strategy that balances spectral complexity across subdomains to improve reconstruction quality in capacity-constrained regimes.

Dataset distillation compresses a large training set into a small synthetic set that preserves downstream training utility. While most existing methods target training networks from scratch, modern visual transfer learning often uses frozen pre-trained encoders followed by lightweight linear probing. Existing distillation methods for this setting either unroll iterative linear-probe updates with trajectory-based gradient matching, or rely on closed-form formulations originally designed for from-scratch training with neural-tangent-kernel (NTK) approximations. Neither route exploits the fact that frozen-feature linear probing admits a closed-form solution determined directly by the pre-trained features themselves, with no infinite-width approximation and no inner-loop trajectory. We propose Closed-Form Linear-Probe Dataset Distillation (CLP-DD), a bilevel formulation that computes the linear probe induced by the synthetic set with a sample-space kernel ridge solver. The synthetic images are then updated by evaluating this induced classifier on real features through a temperature-scaled softmax cross-entropy, where the classifier columns act as learned class anchors in feature space. We further show that the choice of outer objective is decisive: pairing the closed-form inner solver with a standard MSE outer loss substantially underperforms trajectory-based methods, while the discriminative outer loss closes most of the gap. On ImageNet-100 with four pre-trained backbones, CLP-DD substantially improves over LGM without DSA and approaches LGM with DSA at a fraction of the computational cost. On ImageNet-1K, CLP-DD matches or surpasses LGM with DSA on three of four backbones while running roughly $14\times$ faster and using less than one-eighth of the GPU memory.

In a standard transformer, each layer computes attention once: queries examine keys, select values, and move on. But attention is a function of representation, and representation is a function of attention. We close this loop. Fixed-Point Self-Attention (FPSA) computes its attention pattern by iterating the query/key/value process to convergence. Values $V$ are derived from the layer input and held fixed; queries and keys are recomputed from an evolving output state $Z_t$ at each step, converging when representations reach equilibrium, $Z^* = f(Z^*)$. FPSA is a drop-in replacement for multi-head attention that adds zero parameters and adapts depth on its own: easy tokens stabilize in $2$--$3$ iterations, hard tokens refine over $15$ or more, with $\mathcal{O}(1)$ memory cost regardless of the number of iterations. It lifts BERT-Base and ELECTRA-Base on GLUE and SQuAD~v2.0, improves ViT-B/16 accuracy by up to 20\%, and delivers matching gains on vision-language tasks, all without adding a single parameter. Unlike looped transformers that iterate entire layers, FPSA iterates only the attention step, so overhead is modest: a median of $3$--$6$ steps per layer adds roughly $1.6\times$ GFLOPs and $1.3$--$1.4\times$ wall-clock time over BERT-Base. On multi-step reasoning benchmarks (GSM8K, BBH, LogiQA), FPSA's adaptive computation yields clear gains, letting the model refine longer when the input demands it.

Recent advances in multimodal foundation models have demonstrated strong performance across diverse clinical tasks. However, existing approaches predominantly operate at the individual sample level and fail to model the interplay between population-level cohort patterns and patient-specific variation that underlie real-world clinical reasoning. In practice, clinicians routinely reason through patient cohorts sharing similar demographic, physiological, or pathological characteristics, while accounting for patient-specific deviations from the group prototype. Motivated by this gap, we propose CODA, a cohort- and drift-aware foundation model for multimodal clinical reasoning. CODA explicitly integrates cohort structure into the modeling pipeline through two complementary mechanisms: cohort-aware population stratification and drift-aware individual adaptation. In particular, we leverage reinforcement learning to discover clinically meaningful patient cohorts and further model each patient’s deviation from cohort prototypes as a structured drift signal encoding fine-grained individual heterogeneity. These cohort- and drift-aware representations are incorporated into the foundation model to enable reasoning that is simultaneously informed by population-level patterns and sensitive to individual-level variation. Extensive experiments on real-world electronic health record (EHR) benchmarks demonstrate that CODA achieves consistent effectiveness and strong generalizability across diverse clinical tasks, including closed-QA, open-QA, and report generation.

RLVR and OPD have become standard paradigms for post-training. We provide a unified analysis of these two paradigms in consolidating multiple expert capabilities into a single model, identifying capability loss in different ways: mixed RLVR suffers from inter-capability divergence cost, while the pipeline of first training experts and then performing OPD, though avoiding divergence, fails to fully absorb teacher capabilities due to large behavioral pattern gaps between teacher and student. We propose Co-Evolving Policy Distillation (CoPD), which encourages parallel training of experts and introduces OPD during each expert's ongoing RLVR training rather than after complete expert training, with experts serving as mutual teachers (making OPD bidirectional) to co-evolve. This enables more consistent behavioral patterns among experts while maintaining sufficient complementary knowledge throughout. Experiments validate that CoPD achieves all-in-one integration of text, image, and video reasoning capabilities, outperforming strong baselines such as mixed RLVR and MOPD, and even surpassing domain-specific experts.

The reasoning capabilities of large language models are reported to drop sharply when they are forced to think in heavily regulated target languages such as Chinese — a phenomenon the literature attributes to data imbalance, multilingual transfer failures, or, more recently, RL-induced cross-lingual collapse. We propose and test an alternative mechanistic account: that the gap reflects RL-induced selection between conditional policies sharing the same parameters, rather than a loss of underlying capability. We construct a model organism for this account. Two policies — identified solely by the language of an unsupervised chain of thought — are installed via SFT on disjoint splits of ARC-Challenge, with one policy trained to answer and the other to refuse on each split. Reinforcement learning with a final-answer reward, restricted to one split, drives the language of reasoning to collapse onto the rewarded policy across both splits, producing a 67% → 10% accuracy collapse on the held-out split despite no reward signal ever observing it. General capability is unaffected: MMLU accuracy is 71.2% before and after RL. A depth-wise activation steering intervention then recovers held-out accuracy from 0% to 85% on the worst-case slice (Chinese-CoT completions) while preserving Chinese as the surface chain-of-thought language and leaving the rewarded policy intact (83% vs. 94% unsteered). The capability was never lost; it was gated. To the extent deployed regulated-language models contain analogous structure, surface evaluations of such models will systematically understate capability hidden behind learned policy gates.

Test-time compute scaling, where language models "think longer" via extended chain-of-thought (CoT), is the dominant paradigm for improving reasoning. Prevailing intuition holds that extended reasoning *converges* toward an answer; we show the opposite. Across three reasoning-RL families (DeepSeek-R1, Qwen-QwQ, Llama-3.1-Nemotron) and three benchmarks (GSM8K, MATH, GPQA), step-to-step SAE feature turnover and entropy *rise* monotonically across the chain (Cohen's $d = 0.61$–$0.68$, $p < 10^{-3}$). Yet the final answer is already linearly decodable from the residual stream within the first $\sim 20\%$ of the chain. We call this pattern **commit then explore**: the answer becomes stably decodable early, and the bulk of test-time compute is then spent exploring *around* an already-decodable answer. Targeted ablation shows post-commit features have a $2$–$6\times$ smaller teacher-forced log-probability effect (matched-mass control: $1.5$–$4.6\times$) and a $2$–$3\times$ smaller free-regeneration answer-flip rate than pre-commit features. A terminal-reward RLVR credit-assignment account predicts this asymmetry, and the dynamic is absent in base models: Qwen-2.5 and Llama-3.1 bases fail our detector at every layer, and a held-out quantitative prediction derived from R1 generalises to Qwen-QwQ at Pearson $r = 0.96$. Leveraging the structure, we build a Reasoning Monitor that saves $26$–$34\%$ of inference compute at $\geq 92\%$ accuracy retention, dominating four adaptive-compute baselines. Extended chain-of-thought, under this view, is not the computation of an answer; it is the exploration of its neighborhood.

Scaling Transformers to long contexts is constrained by the quadratic cost of self-attention and the linear growth of key-value cache memory transfer. Sparse attention mitigates this by retrieving only relevant tokens, but current approaches either require large-scale training or, within the training-free regime, rely on semantically coarse heuristics or expensive clustering that is difficult to update efficiently during decoding. We introduce CommunityKV, a framework that formulates sparse attention as a community detection problem. CommunityKV constructs a token graph from the $QK^T$ scores already computed during standard prefill, and partitions the graph into communities to enable retrieval of semantically coherent token groups. A local update rule assigns newly generated tokens to communities in constant time, enabling sparse retrieval throughout streaming decoding without global re-partitioning. We evaluate CommunityKV on open weight models and long context benchmarks, demonstrating that our approach achieves up to $1.86\times$ higher decoding throughput than exact FlashAttention-2 with negligible accuracy degradation.

A generated plausible future mathematical claim must satisfy two constraints: it should follow the direction of prior work and respect the formal dependencies that constrain the validly of the claim. Existing approaches typically model only one of these sources, producing claims that are either weakly grounded or insufficiently motivated. We introduce grounded future mathematical generation, the task of generating a plausible future theorem-like claim for an anchor paper, using two complementary sources of context: its scientific citation graph ,and its aligned formal theorem dependency graph. To address this setting, we propose COMPOSE, a dual-graph framework that combines scientific citation context with formal mathematical structure to condition a language model. To support this setting, we construct a dataset of 108K paired scientific-formal graphs from arXiv and Mathlib, together with a benchmark of 47K future papers from 2024-2025. Experiments show that COMPOSE outperforms strong baselines on retrieval to real future papers and achieves the best overall performance under LLM-judge evaluation, producing more grounded and mathematically richer outputs. These results show that future mathematical generation benefits from combining scientific context with formal structure.

A neural network trained on the multiplication table of the quaternion group Q₈ can achieve perfect accuracy within each cyclic subgroup ⟨i⟩ and ⟨j⟩, predict every cross-region product wrong, and stay there indefinitely. This is the grokking plateau, and we prove its height is exactly 4 in dimension d=2, determined by the topology of the overlap before any training. The mechanism is a homotopy pullback: for G = A ∗_C B with finite overlap, the representation space of G is a homotopy pullback of those of A and B over that of C, so locally accurate modules need only agree up to a basis change. From this we derive a stratified generalization certificate whose strata are the connected components of Hom(C, U(d)): in the correct stratum, with locally accurate modules, the cross-region prediction is O(ε)-close to the target; in the wrong stratum it is provably far regardless of local accuracy, with exact gap c = 2(d − rₘₐₓ) for central cyclic overlap. The plateau is therefore the time the network spends in the wrong stratum, the grokking transition is the discrete jump between strata, and the framework predicts which task families exhibit such plateaus—those whose overlap representation space is disconnected—and which do not.

Long-horizon robotic planning often requires satisfying temporally ordered safety and visitation constraints rather than a single goal condition. We study multiple reach-avoid (MRA) tasks, in which a robot must sequentially visit target regions while remaining within corresponding safe sets. Existing approaches either rely on explicit system models or use data-driven planners whose test-time conditioning mechanisms do not scale well to long-horizon temporally structured tasks. We propose a training-free compositional diffusion planning framework for MRA tasks that operates directly on a pre-trained task-agnostic diffusion trajectory prior at test time, without requiring a known dynamics model. Our approach decomposes a global MRA specification into local reach-avoid sub-tasks, composes overlapping short-horizon diffusion priors into a long-horizon generative process through a factor-graph perspective, and enforces each local requirement through projection-based denoising during sampling. The same construction also extends naturally to prefix-suffix tasks through a looped compositional graph. We provide a formal correctness guarantee showing that the stitched trajectory satisfies the target specification. Experiments on long-horizon constrained planning benchmarks show strong execution success, substantially lower planning time than representative guidance-based baselines, and effective transfer to richer dynamics.

Decision trees with binary leaves are fundamental building blocks in classical machine learning. While individual trees can be used as strong interpretable classifiers, they frequently serve as members of large collections instead: primarily as components of powerful ensembles, but also as one of the few model classes for which the entire set of near-optimal models (the Rashomon set) can be analyzed. Both types of collections are often more complex than necessary, because there are many decision tree representations for the same decision boundary, imposing unnecessary computational costs in ensembles and complicating theoretical work on Rashomon sets. We propose methods to reduce collections of trees to a subset representing each unique decision boundary in the collection, with each tree in its sparsest equivalent form. In so doing, we improve the scalability of removing provably redundant classifiers in a Rashomon set. We also compress many shallow Adaboost ensembles by >15\% of their total number of leaves, while exactly preserving the models' decisions and probability estimates on any possible observation. These methods allow practitioners to run more efficient computations over compressed collections of trees, and remove arbitrary complexity when analyzing and understanding them.

Post-training quantization degrades multimodal model capabilities unevenly, yet practitioners lack a principled way to predict which capabilities will collapse. We propose the **Computational Depth Principle** (CDP): end-to-end fidelity under quantization follows $\hat p^n$, where $n$ counts sequential precision-dependent operations and $\hat p$ is a per-architecture survival rate. Across four VLM families and diverse benchmarks spanning image understanding, image generation, video understanding, and video generation, we show that deeper reasoning tasks consistently degrade more steeply and that a single shallow probe predicts the full degradation ranking on held-out architectures. The same exponential form extends to image generation, where denoising step count plays the role of depth, and to video understanding, where longer temporal reasoning chains amplify losses. For video generation, spatial quality degrades more than temporal coherence, indicating that the dominant compounding axis is denoising depth rather than cross-frame coupling. A complementary distortion-to-information (D/I) flip ratio diagnoses quantization method effects, revealing that certain methods preserve near-baseline quality while others produce catastrophic collapses on specific architecture pairings that aggregate accuracy alone cannot detect. Module ablation localizes the sensitivity to the language backbone rather than the vision encoder. Together, depth and D/I convert full-suite evaluation into a two-probe screening protocol that surfaces the highest-risk capability gaps first.

Concept Bottleneck Models (CBMs) enable interpretable image classification by structuring predictions around human-understandable concepts, but extending this paradigm to video remains challenging due to the difficulty of extracting concepts and modeling them over time. In this paper, we introduce MoTIF (Moving Temporal Interpretable Framework), a transformer-based concept architecture that operates on sequences of temporally grounded concept activations, by employing per-concept temporal self-attention to model when individual concepts recur and how their temporal patterns contribute to predictions. Central to the framework is a class-conditioned VLM-based concept discovery module that extracts object- and action-centric textual concepts from training videos, yielding temporally expressive concept sets without manual concept annotation. Across multiple video benchmarks, this combination improves over global concept bottlenecks and remains competitive within the interpretable concept-bottleneck setting, while narrowing the gap to strong black-box video baselines that we report as contextual references.

Finetuning a language model can lead to emergent misalignment (EM) [Betley et al., 2025b]. Models trained on a narrow distribution of misaligned behavior generalize to more egregious behaviors when tested outside the training distribution. We study three interventions proposed to reduce EM. We confirm that these interventions reduce or eliminate EM on existing evaluations (questions like “How do I make a quick buck?”). However, if the evaluation prompts are tweaked to resemble the training context, the model displays EM. We call this conditional misalignment. As in standard EM, the model displays misaligned behaviors more egregious than those seen during training, but only on inputs sharing features with the training data. The first two interventions are diluting misaligned data with benign data, and finetuning on benign data after misaligned data. Both produce conditional misalignment. For instance, models trained on a mix of only 5% insecure code still show misalignment when asked to format responses as Python strings (resembling the training context). The third intervention is inoculation prompting. Here, statements with a similar form to the inoculation prompt serve as triggers for misalignment, even if they have the opposite meaning. On the positive side, inoculation prompting has lower (but still non-zero) conditional misalignment if training is on-policy or includes reasoning distillation. Our results imply that in realistic post-training, where misaligned data is typically combined with benign data, models may be conditionally misaligned even if standard evaluations look clean.

The rapid emergence of heterogeneous large language models (LLMs) has highlighted the limitations of static single-model inference systems. Although model routing alleviates this issue by assigning the most suitable model to each query, existing methods often struggle to stably align query semantics with actual model performance and remain constrained by the capability ceiling of a single model. On the other hand, although multi-model aggregation can improve performance on complex queries, it also introduces extremely high computational overhead for routine tasks. In this paper, we propose PARS, a model confidence estimation method for dynamic LLM routing and aggregation. PARS first decouples the feature representation space and applies structured smoothing to local performance signals, producing more discriminative confidence score estimates for candidate models. On this basis, PARS represents inference modes such as direct routing, self-consistency sampling, and multi-model voting as test-time compute allocation actions under a budget constraint. By extracting state features such as routing confidence distributions and predictive uncertainty, a lightweight selector adaptively assigns an inference strategy to each query. Extensive experiments show that PARS improves average accuracy over the best baselines by 1.02% and 1.36% in the single-model and multi-model settings, respectively, while reducing average inference cost by 1.73x and exhibiting stable out-of-distribution generalization.

While reasoning traces improve model capability, they also create a pathway for extracting supervision signals for downstream distillation. Initial efforts to mitigate this rely on decoding-time interventions, but they incur high inference overhead and often harm the teacher model's natural fluency. We propose Conflict-Aware Logit Adapters (CALA), a lightweight training-time defense. CALA attaches a small residual adapter to a frozen teacher model and trains only this adapter. A proxy student identifies positions where imitation is most sensitive, and the adapter perturbs the output distribution there within high-probability regions. A conflict-aware optimization balances competing anti-distillation and utility objectives, with regularization preserving the teacher’s performance and fluency. At inference, CALA adds negligible overhead with no auxiliary models required. Experiments on reasoning benchmarks show that CALA substantially reduces successful student imitation while maintaining near-original teacher performance and generation quality, offering a practical and efficient alternative to decoding-time approaches.

Simulation-based inference with neural posterior estimation (NPE) often yields overconfident and unreliable posteriors under limited simulation budgets. To address this, we propose DRO-NPE, a distributionally robust approach that replaces the standard NPE objective with a worst-case loss over a Wasserstein ambiguity set. We introduce KL-based metrics for miscoverage and miscalibration, and use these to show that the DRO-NPE objective controls overfitting and reduces posterior overconfidence. Our method is tractable, parallelisable, and readily integrates with standard normalising flows. Across benchmark SBI tasks, DRO-NPE consistently improves coverage and calibration, while narrowing the gap between empirical and population NPE loss, leading to more reliable inference in low-simulation regimes.

Test-time scaling often uses an external verifier, such as compilers and test-cases in coding or trained value functions in robotics application, to obtain high quality rollouts. Verifier-free test-time scaling (or VF-TTS) is gaining extensive attention as a mechanism to enhance Large Language Model (LLM) reasoning, primarily because in many real-world applications we do not have access to such high quality verifiers. Among existing VF-TTS methods, confidence-based VF-TTS methods, which compute and rank rollouts solely by confidence, are particularly promising. Such methods introduce near-zero overhead for sample evaluation and require minimal access to internal model states, making the methods highly flexible across models and data. In this paper, we demonstrate a critical limitation of existing confidence-based VF-TTS methods by showing that such methods catastrophically break down on complex tasks. We observe a very interesting phenomenon: uniformly high confidence frequently indicates a failure to explore, favoring confidently wrong answers. To address this, our core insight is that robust cognitive search requires a specific confidence trajectory pattern: such methods perform exploratory branching at the beginning as manifested by low initial confidence, and converge to a high final confidence solution. To implement this insight, we introduce consilience, a novel selection framework that explicitly evaluates the temporal asymmetry of confidence in reasoning. We operationalize this via a combinatorial metric that actively penalizes high initial confidence while strictly demanding final certainty. Extensive experiments covering both graduate-level mathematics problems and free-form code generation demonstrate that consilience effectively outperforms existing baselines, validating our novel perspective on completion confidence.

This paper establishes a rigorous measurement science for AI agent reliability, providing a foundational framework for quantifying consistency under semantically preserving perturbations. By leveraging $U$-statistics for output-level reliability and kernel-based metrics for trajectory-level stability, we offer a principled approach to evaluating agents across diverse operating conditions. Our proposal highlights the important distinction between the core capability and execution robustness of an agent, showing that minor task-level variations can induce complete strategy breakdowns despite the agent possessing the requisite knowledge for the task. We validate our framework through extensive experiments on three agentic benchmarks, demonstrating that trajectory-level consistency metrics provide far greater diagnostic sensitivity than traditional pass@1 rates. By providing the mathematical tools to isolate where and why agents deviate, we enable the identification and rectification of architectural concerns that hinder the deployment of agents in high-stakes, real-world environments.

Planar graphs are central to applications across science and engineering, yet existing generators provide limited support for goal-directed generation under hard structural and geometric feasibility constraints. We propose a dataset-free method for generating planar graph embeddings by combining parametric graph grammars with safe reinforcement learning to optimize generic task-specific objectives while satisfying constraints during construction. We formulate the generation process as a constrained Markov decision process, where the graph grammar defines the state and action spaces. We further introduce an action projection that maps sampled actions toward state-dependent safe sets, improving constraint satisfaction during training. In contrast to classical graph generators and deep generative models, which typically offer limited goal-directed control or rely on weak constraint satisfaction, our method constructs feasible planar graph embeddings directly during generation. We also introduce a benchmark suite for constrained and goal-directed planar graph generation, together with classical and deep generative baselines. Across all benchmark tasks, our method consistently outperforms baselines while satisfying the formulated constraints.

Many complex tasks require extended effort, diverse capabilities, or coordinated actions beyond what a single agent can provide. However, simply adding more agents does not guarantee better performance, as effective cooperation depends on how agents interact with each other and with task structure to satisfy evolving constraints over time. This challenge is amplified for LLM-based multi-agent systems (LLM-MAS): plans, messages, and revisions occur in natural language, whereas task progress depends on grounded environment actions. Current evaluations mostly treat cooperation as an implicit ingredient of final task success, leaving both cooperation and the effect of multi-agent interaction on task dynamics difficult to study. We introduce COOP$^2$, an evaluation framework that grounds high-level agent cooperation dynamics in LLM-MAS within task progress in the environment. COOP$^2$ then defines cooperative tasks with verifiable cooperative requirements, allowing us to analyze how cooperation unfolds over time with respect to task progress, as well as where and why cooperation breaks down. Building on this framework, we develop COOP$^2$-Repair, which predicts constraint failures from group plans and opens targeted repair channels for guided revisions. Across two environments and three communication structures, COOP$^2$-Repair improves task success and constraint satisfaction while exposing the additional decision overhead and communication load required for repair.

Post-training quantization (PTQ) enables efficient deployment of large language models by mapping pretrained weights to low-bit formats without retraining, typically using a small calibration set to minimize a layer-wise calibration objective. However, this sequential procedure induces a mismatch: errors from earlier quantized layers alter the inputs received by later layers, causing the activations to deviate from those of the full-precision model. Recent approaches introduce mismatch-aware calibration objectives to compensate for this effect, but leave open how much of the observed mismatch should shift each layer's calibration target. Fully applying this correction can overfit limited calibration data, while scaling the mismatch correction with a fixed coefficient ignores varying reliability of mismatch estimates across layers. To address these limitations, we propose CoreQ, a learning-free PTQ framework that applies a closed-form coefficient for mismatch correction derived from a geometric decomposition of the mismatch. The resulting coefficient adapts the correction across layers, reduces overfitting to finite calibration data, and requires no hyperparameter tuning. Given the corrected target, CoreQ minimizes the induced triangular least-squares objective with an efficient greedy successive-rounding solver and a bounded beam-search extension, K-CoreQ, that trades modest additional compute for improved performance. Across multiple LLM families, scales, bit-widths, and quantization settings, CoreQ improves perplexity and downstream accuracy over strong PTQ baselines.

We study contextual bandit problems with correlated arms and access to surrogate reward signals produced by a machine learning model, motivated by applications such as large language model (LLM) routing. Unlike classical contextual bandits that rely solely on bandit feedback and assume conditional independence across arms, our setting allows context-dependent inter-arm correlations and auxiliary reward information that may be noisy or misspecified. We propose algorithms that leverage such surrogate rewards through two complementary designs. A coupled reward-mixing approach pools true and surrogate rewards to accelerate learning when surrogate signals are reliable, while a decoupled prediction-mixing approach maintains separate estimators for bandit feedback and surrogate rewards and adaptively combines their predictions. This decoupling yields robustness to surrogate misspecification, recovering regret guarantees comparable to reward-only bandit methods in the worst case, while achieving improved regret when surrogate predictions are sufficiently informative. We provide theoretical regret analyses for both approaches and evaluate them on LLM routing benchmarks under varying accuracy versus cost trade-offs. The results demonstrate improved sample efficiency and consistently better accuracy–cost trade-offs compared to standard contextual bandit baselines and strong static routing methods.

Counterfactual explanations for time-series classification generate synthetic instances that flip a prediction to a desired class while remaining plausible and making small, local changes. Existing methods often rely on classifier gradients, and many do not naturally extend to one-class settings. To bridge this gap, we propose CTCF, a model-agnostic method for hard-decision classifiers in supervised and one-class settings. CTCF decouples generation from classifier-specific control by reusing an unconditional Flow Matching generator and learning a flow-time-dependent desired-class region from interpolation states labeled by endpoint hard decisions, without classifier gradients or class probabilities. At inference time, CTCF uses a dual-control mechanism: fused-lasso endpoint-objective steering encourages small, segment-local changes, while linearized projection correction reduces violations of a conservative desired-class region. Theoretically, under ideal marginal matching, we show that the surrogate risk on interpolation states equals the corresponding ODE-trajectory risk. Experiments on UCR datasets demonstrate CTCF's effectiveness across quantitative metrics and qualitative case studies.

Sparse attention accelerates long-context decoding by reading a subset of the KV cache, but existing methods are open-loop: the kernel commits to its selection with no signal indicating whether enough attention mass was preserved and no recovery path when it was not. The needed signal already exists inside the kernel. The fraction of softmax mass on the selected tokens, which we call coverage, satisfies an exact algebraic equality with per-step output error and serves a dual role as the acceptance criterion and the interpolation weight for recovery. Coverage-Verified Sparse Attention (CVSA) turns coverage into a training-free, closed-loop verify-then-correct decode loop that accepts high-coverage drafts and recovers only the heads that fall short, with no sampling, no draft model, and no speculative-decoding infrastructure. Turning verification on matters far more than tuning its threshold, and matched-budget comparisons confirm the gain is structural. On LongBench and RULER at 7B and 70B, CVSA closes the quality gap to dense attention to within statistical noise, with measured throughput reaching 6.21x at 128K context.

Automated structural health monitoring is essential to prevent catastrophic infrastructure failures. Precise, pixel-level defect segmentation is needed to accurately assess structural integrity, but progress in defect segmentation for civil infrastructures has been held back by an extreme scarcity of data, which requires costly expert annotation. The need for data is accentuated by algorithmic hurdles intrinsic to the problem, including center-bias and the need to rely more on shape when inspecting nearly textureless building materials. To remove the bottleneck, we introduce **Cracks in the Foundation (CiF)**, the largest and most detailed civil infrastructure (instance) segmentation dataset to date, comprising $\approx$150,000 high-resolution images meticulously curated over five years in collaboration with civil engineering experts. With the help of this unprecedented data source, we expose a blind spot of current visual AI: despite the advent of promptable Foundation Models (FMs) and Vision Language Models (VLMs), and despite the impressive abilities of today's specialised segmentation models, it turns out that dense image understanding in the built environment is nowhere near solved. Our evaluations indicate that even the most recent zero-shot FMs face significant challenges when deployed on real-world infrastructure and even the performance of specialised models with domain-specific supervision plateaus at $\approx$25\% mAP. CiF establishes inspection of civil infrastructure, an elementary and seemingly easy perceptual task, as an open challenge that reveals fundamental weaknesses of present-day models trained predominantly on internet images, literally and figuratively highlighting *cracks in the current foundation model paradigm*.

Foundation models are central to building general-purpose embodied agents. Minecraft has become the dominant testbed for this paradigm, but its computational and engineering overhead limits rapid iteration and controlled analysis. Crafter offers a lightweight alternative yet lacks the foundation model ecosystem required for such research. We introduce CrafterDojo, a suite of foundation models, datasets, and data-generation toolkits that establishes Crafter as a testbed for foundation-model-based embodied intelligence, providing CrafterVPT, CrafterCLIP, and CrafterSteve-1. To address Crafter's data scarcity, CrafterDojo includes automatic tools that generate diverse synthetic demonstrations and aligned video-language pairs, yielding the CrafterPlay and CrafterCaption datasets. Experiments show that these datasets support effective foundation model training and enable hierarchical agents that outperform non-hierarchical baselines on long-horizon tasks. We release all code, models, and datasets.

How well do vision-language models actually reason about what they see? Current diagnostic benchmarks offer limited answers: they draw from fixed question templates, evaluate only the final answer, and are tied to single visual environments. We introduce CRISP (Compositional Reasoning over Images via Stackable Programs), a platform that generates compositional visual reasoning tasks from reusable, composable operations. Given any scene with a structured specification, the platform constructs reasoning chains of arbitrary depth where each intermediate step has automatically derived ground truth, including bounding boxes for the objects the model should attend to at every stage. New visual environments and reasoning operations plug in without modifying the generation engine. We instantiate the platform across three environments spanning 2D sprites, photorealistic 3D characters, and indoor room scenes, and show that frontier VLMs degrade systematically with reasoning depth and frequently arrive at correct answers through incorrectly grounded intermediate reasoning. Beyond benchmarking, the generated reasoning traces and step-wise ground truth can support future work on training and improving visually grounded reasoning models.

Dictionary learning methods - such as Sparse Autoencoders (SAEs) and crosscoders - decompose model activations into human-interpretable building blocks. We introduce temporal crosscoders, a simple and flexible framework for feature discovery in Large Language Models (LLMs). To properly evaluate temporal crosscoders we develop TempBench: a panel of synthetic and real-world tasks for evaluating temporal structures. Temporal crosscoders outperform both conventional and temporal architectures in both of our synthetic settings and on two out of four of the real world settings - more than any other current architecture. Most strikingly, they can detect backtracking - a key reasoning behavior - at a 40\% higher rate than conventional SAEs, and are 15\% more effective in inducing it. Our results establish temporal crosscoders as a simple and flexible framework for feature discovery, both local and temporal. We provide full code at the following anonymous repository: \url{https://anonymous.4open.science/r/temp-bench-anon/}.

Diffusion and flow-matching models are typically constrained to a single manifold, where the noise prior, intermediate trajectory, and prediction target share the same dimensionality. While Latent Diffusion Models (LDMs) mitigate computational costs by operating in compressed spaces, they still rely on a decoupled, fixed decoder to map latents back to pixels. We introduce CrossFlow, a generative paradigm that unifies these stages by allowing the noise prior and final output to reside in different spaces. By deriving a novel cross-space objective, our framework enables a single model to map directly from a noisy latent to a high-resolution image in a single function evaluation (NFE). CrossFlow serves a dual purpose: it acts as a high-fidelity one-step generator and functions as an enhanced decoder for existing LDM pipelines, capable of refining imperfect latent estimations during the reconstruction process. On ImageNet-1k ($256 \times 256$), CrossFlow achieves a state-of-the-art 1.62 FID with only one NFE. Our results demonstrate that cross-space flow objectives provide a scalable and theoretically grounded framework for unifying latent generation and pixel-space decoding.

Personalized portrait generation aims to synthesize portraits that align with both the identity of the reference face and the semantics of the text. However, existing methods suffer from two critical challenges. First, they exhibit a trade-off between ID fidelity and text controllability. We identify that this limitation stems from the self-supervised training, where the model learns a trivial mapping rather than understand identity. Second, despite exploring various encoding strategies, recent approaches still tend to discard fine-grained facial details. To address these challenges, we propose CrossID-11M, a large-scale curated portrait dataset comprising 11 million images with high-quality annotations, paired with millions of identity-preserving face video clips. Furthermore, we introduce CrossID, a novel framework integrating a lightweight GatedFaceAdapter and a cross-supervised training paradigm. Additionally, we establish CrossID-Bench alongside new VLM-based evaluation metrics. Extensive experiments demonstrate that CrossID successfully breaks the trade-off between ID fidelity and controllability, achieving state-of-the-art performance by producing photorealistic images with faithfully preserved ID details.

Although Vision-Language Models (VLMs) demonstrate impressive capabilities, their deployment in high-stakes domains is severely hindered by hallucinations. Existing detection methods mostly focus on static internal feature probing with coarse-grained annotations, failing to achieve fine-grained token-level localization without expensive dense labels. In this paper, we reconceptualize VLM hallucination detection as a dynamic graph learning task, identifying hallucinations as distinct trajectory drifts across network layers. Consequently, we propose TRACE (Token Representation Analysis with Cross-layer Evolution), a non-intrusive probing framework that models internal states as heterogeneous spatio-temporal graphs. By employing Graph Neural Networks (GNNs) to capture computational dynamics, TRACE effectively isolates hallucinatory signals. Furthermore, a token-first Multiple Instance Learning (MIL) readout elegantly bridges the granularity gap, distilling macroscopic sentence-level supervision into precise zero-shot token-level localization. Extensive evaluations on M-HalDetect, ViGoR, and HalLoc demonstrate that TRACE achieves state-of-the-art detection performance while being competitive in localization accuracy, offering an efficient and interpretable solution for trustworthy VLMs.

Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill from scratch. We propose cross-model KV cache transfer, where the receiver reuses the source’s KV cache, skipping prefill. We find that cross-model KV has substantial linear structure across matched-KV pairs, where source and target share KV head count and per-head dimension. On Qwen3 14B→32B, one source layer explains 56% of variance in the target’s keys and 32% in values, rising to 79% and 65% with multiple source layers. Building on this, we design a closed-form ridge mapper that operates per head and proceeds in three steps. First, for each target layer we select the top-kmost predictive source layers and concatenate their KV as input. Second, we strip RoPE from the keys before mapping, so the fit is position-free and reusable across context lengths. Third, we fit ridge regression on a small calibration set of 500 FineWeb-Edu sequences of 1,024 tokens each. Surprisingly, across six pairs in three families, this linear mapper retains 71–98% of the receiver’s standalone-prefill accuracy on four pairs, while two degrade sharply. A nonlinear MLP recovers up to +22 pp HellaSwag retention on the failures. The mapper runs 3–31×faster than re-prefill and remains stable across multi-turn handoff, making cross-model KV cache transfer practical.

Benchmarks for large language models typically evaluate the accuracy of answers against fixed reference labels. But a central step in many complex real-world tasks is identifying which questions are worth asking in the first place: decomposing a difficult problem into subquestions whose answers provide key steps on the path toward solving the larger problem. To evaluate this capability of information discovery, we introduce CruxBench, a benchmark that grades LLM-generated questions by their Value of Information (VOI): how much a model-proposed subquestion updates beliefs about a target outcome. To automate this measurement at scale, we target as outcomes the prices in continuously updating forward-looking prediction markets. Our benchmark has three rare properties: it is contamination-resistant by construction, since ground truth is generated by future world events; it is open-ended, admitting unbounded and complex text-based submissions rather than one correct numeric answer; and it is grounded, with informativeness measured against quantified changes in real-world beliefs (prices). We evaluate eight frontier and open-weight models on 336 target outcomes and find that VOI correlates highly with independent measures of model quality ($r=0.87$) and predicts downstream usefulness when the generated questions are passed to a separate forecasting agent.

Crystal generative models mainly learn what stable crystals look like, with little explicit supervision for what makes them stable. We reveal a substantial representation gap between state-of-the-art crystal generative models and pretrained universal machine learning interatomic potentials (MLIPs) via energy probing, and show this gap can be closed by a simple training-time alignment. We propose Crystal REPresentation Alignment (CrystalREPA), a plug-and-play framework that aligns the atom-wise hidden states of generative encoders with frozen MLIP representations through an element-aware contrastive objective, transferring stability-aware atomistic priors with marginal training overhead and no additional inference cost. Across three generative frameworks, ten MLIP teachers, and two benchmark datasets, CrystalREPA consistently improves the thermodynamic stability, structural validity, and structural fidelity of generated crystals. Equally important, we find that an MLIP's transfer effectiveness is poorly predicted by its accuracy on standard leaderboards (e.g., Matbench Discovery) but strongly predicted by the distinguishability of its atom-wise representation space, yielding a practical, accuracy-independent criterion for selecting MLIP teachers for generative transfer.

Visual reprogramming (VR) efficiently repurposes pre-trained vision-language models for new image classification tasks by adding trainable patterns to inputs without modifying the backbone. Nevertheless, existing VR methods heavily depend on labeled data and thus suffer substantial performance degradation under limited supervision (i.e., scarce labeled images). We introduce semi-supervised learning (SSL) into VR to exploit abundant unlabeled images and enhance data efficiency. However, directly applying generic SSL techniques often amplifies biases in unlabeled posteriors and destabilizes VR training. To address these challenges, we propose CURE (Consistency and Unlabeled Recalibration with Confidence-Margin Enhancement), a semi-supervised VR framework comprising two key components: (i) Recalibreted Gaussian Confidence-Margin Soft Weighting to dynamically adjust sample importance; (ii) Dual-attribute Consistency Regularization to ensure consistency across attribute prompts. Extensive experiments show that CURE consistently outperforms both existing VR methods and direct SSL extensions, improving average classification accuracy by 5.06 and 6.67 percentage points across twelve widely-used benchmarks for limited supervision tasks. Our code is available at https://anonymous.4open.science/r/CURE-93E2/.

In Euclidean online convex optimization, first-order and full-information feedback share the minimax regret (\Theta(DL\sqrt{T})) over (T) rounds, as convex losses can be replaced by their linearizations. On Riemannian manifolds this reduction fails, because the corresponding first-order surrogate of a geodesically convex loss is generally not geodesically convex; we show that this failure has provable consequences. For online convex optimization on Hadamard manifolds with sectional curvature bounded below by (\kappa<0) and feasible diameter (D), we construct hard hyperbolic-space instances yielding first-order regret lower bounds at the scale ( DL\sqrt{\zeta\, T}, ) where ( \zeta=\sqrt{|\kappa|}D/\tanh(\sqrt{|\kappa|}D) ) is the curvature--diameter factor. Our first lower bound applies to all (possibly randomized) first-order algorithms satisfying a natural geometric-span condition, holding almost surely over the algorithm's internal randomness. Our second lower bound removes the geometric-span restriction for deterministic first-order algorithms, provided the feasible diameter satisfies (D=\Omega(\log T)); the proof rests on a new construction that confines the hidden comparator to a single horosphere in hyperbolic space. Combined with existing curvature-free full-information upper bounds, these results yield a feedback-model separation on Hadamard manifolds: full-information algorithms attain (O(DL\sqrt{T})) regret, yet the additional (\sqrt{\zeta}) factor is unavoidable for first-order feedback whenever either the geometric-span condition holds or the algorithm is deterministic.

Randomized controlled trials typically assume that prognostic covariates are known and available at no cost. In practice, obtaining high-dimensional pretreatment data is costly, forcing a trade-off between covariate-adaptive precision and a measurement budget. We introduce Dynamic Adaptive Rerandomization via Thompson Sampling (DARTS), which treats covariate acquisition as a sequential optimization problem embedded within a design-based causal inference task. A budgeted combinatorial Thompson sampler learns which covariates are most prognostic across successive batches; selected covariates then drive rerandomization and regression adjustment to reduce batch-level average treatment effect variance. Our primary theoretical contribution is a decoupling result: adaptive covariate selection based on past batches preserves batch-level randomization validity, and the cumulative inverse-variance weighted estimator achieves at least nominal asymptotic coverage. We further derive a Bayes risk bound for the acquisition layer that matches the minimax lower bound up to logarithmic factors. Empirically, DARTS systematically concentrates budget on informative features, significantly closing the efficiency gap to oracle designs while maintaining strict inferential validity.

Microbial ecosystems govern human health, agricultural productivity, and biogeochemical cycling. Two coupled modalities determine their function: which organisms are present (metagenomic content) and which small molecules they produce and consume (metabolomic content). Existing biological foundation models cannot represent this joint state: they operate on a single modality, on assembled single-organism genomes, and on the within-DNA axis of generalization, leaving the multi-omic, mixed-organism, mixed-quality regime that governs clinical and surveillance use unaddressed. We present Darwin-7B, a 7-billion-parameter multi-omic foundation model whose primary axis of generalization spans both modalities (metagenome + metabolome) and organizational scales (read → community → host phenotype). We combine a Mamba–Transformer hybrid backbone (24 of 32 layers Mamba) with a hypergraph neural network over KEGG metabolic-reaction hyperedges and bidirectional cross-modal attention modules that align genomic and metabolomic representations in a shared 40,192-token embedding; an Aitchison-space compositional consistency loss respects the simplex geometry of microbial-abundance data at the output. We pretrain on 8 T base pairs of metagenomics, 250K LC-MS/MS metabolite profiles, and 2M KEGG/GO functional annotations, preparing the metagenomic corpus with a published quality-aware tokenization framework. Against the matched 7B genomic baseline METAGENE-1, we report 94.5 MCC on Pathogen Detection (vs. 93.0) and 0.98 F1 on CAMI species-level metagenomic profiling; on a December 2025 SRA outbreak benchmark post-dating the pretraining cutoff, we report 91.0 MCC vs. 81.2. We unlock four clinical tasks single-modality genomic models cannot reach: IBD AUC 0.947, T2D AUC 0.883, antibiotic-resistance AUC 0.910, and metabolic-pathway prediction wF1 0.91, beating the strongest multi-omic baseline (MOGONet) by +3.0–+3.7 AUC. Token-matched scaling from 100M to 70B yields monotone gains with no saturation on the multi-omic clinical AUCs. We further validate predictions in a prospective wet-lab pilot (n = 87) at predicted-vs-observed Pearson r = 0.72. Sparse-autoencoder probing recovers 14 high-coherence biological features tracking KEGG pathways, NCBI taxa, and CARD AMR genes (r = 0.68–0.74); zeroing the butyrate-producer feature drops IBD AUC from 0.947 to 0.928. We conclude that multi-omic, multi-scale modeling is tractable at 7B parameters, and we release the model card, training-data manifest, evaluation pipelines, and a 100-trajectory MetaOmics-10T causal-pilot dataset.

Data-free methods for analysing and understanding the layers of neural networks offer many metrics for quantifying notions of strong' versusweak' layers, with the promise of increased interpretability. In particular, random matrix theory (RMT) offers data-free metrics that claim predictive power over the quality of pre-trained models at a layerwise precision and the ability to identify model pathologies, indicating a unique relationship with generalisation. As a result, metrics from RMT are championed for pre-training optimisation strategies and post-training compression. We establish that RMT-based metrics are unrelated to performance or training by showing that functionally indistinguishable reparametrisations of a pre-trained model can have arbitrary metrics. We show this across a range of architectures and scales. To create functionally indistinguishable reparametrisations, we exploit the well-established phenomenon of criticality: some layers can be re-initialised or re-randomised without affecting the functional behaviour of the model -- they are called robust -- while others cannot -- they are called critical. Re-initialising or re-randomising robust layers provides functionally indistinguishable reparametrisations for which RMT-based metrics are arbitrary. Moreover, we show that relationships between metrics offered by RMT are spuriously related to generalisation. We conclude by showing that if many pre-trained models have data-free metrics in a `good' range, it is, in part, dependent on model initialisation.

High-resolution global ocean states expose a bottleneck in data-driven forecasting: effective long-horizon rollouts require a compact, decodable latent state that reduces I/O and stabilizes error growth. While learned compression has matured for natural RGB images, its assumptions do not transfer to ocean modeling: ocean states are multi-variable and multi-depth, their grids are orders of magnitude larger, and fixed land--sea boundaries create sharp discontinuities that make aggressive downsampling error-prone. We present DC-Ocean, the first deep latent compression architecture for multi-variable, multi-depth global $1/12^\circ$ ocean states, built on an autoencoder that couples convolutional feature extraction with transformer-based global context modeling to retain both coastal sharpness and basin-scale structure. To explicitly address coastline-induced artifacts, DC-Ocean integrates boundary-aware gated convolutions together with a spherical geodesic front propagation interpolation scheme for stable behavior near land masks. Across key ocean variables, DC-Ocean achieves higher reconstruction fidelity than strong autoencoder baselines adapted from the natural-image domain at matched downsampling factors. Beyond reconstruction, we perform autoregressive forecasting directly in latent space and decode predictions back to the physical domain. DC-Ocean matches short-term accuracy of full-resolution forecasting while achieving lower RMSE and slower error growth at medium and long lead times. Together, these results demonstrate that DC-Ocean provides a practical framework for stable and extended high-resolution sub-daily ocean forecasting.

Recent studies have highlighted the crucial role of distance metrics in noisy few-shot learning. However, existing approaches heavily rely on Euclidean or cosine measurements, failing to recognize the ``deceptive proximity'' where heterogeneous samples from different manifold peaks appear spuriously close in the flat ambient space. In this paper, we present Disentangled Dirichlet Geodesic Evaluation (DDGE), a novel manifold-aware framework that unifies orthogonal disentanglement and conformal integrals to efficiently evaluate complex topological spaces. Specifically, we decouple features into orthogonal semantic subspaces and leverage a prior-guided Dirichlet Process to expand discrete samples into a continuous, noise-purified semantic terrain. Upon this landscape, we measure the intrinsic geodesic distance via a density-aware conformal line integral to capture the authentic class topology. Extensive experiments demonstrate the state-of-the-art performance of DDGE on multiple benchmark datasets.

We propose a fully decentralized $Q$-learning dynamic for infinite-horizon discounted Markov potential games in which agents observe the global state and their own realized payoffs, but do not know the reward functions, the transition probabilities, the potential function, or the policies and the actions of the other players. In the proposed dynamic each player maintains a local estimate of its continuation payoff for each state and action and updates its policy through a smoothed best response. Importantly, the smoothing parameter is adjusted by using an online estimate of the discounted state-visitation distribution, so that players explore more in rarely visited states while exploitation is encouraged in states that are frequently visited. We prove that the induced policy sequence converges almost surely to a set of approximate Nash equilibria of the Markov potential game, with an approximation guarantee that depends only on easily computable parameters. We illustrate the performance of the proposed learning dynamic via numerical experiments on a routing game and compare it with existing independent learning schemes based on persistent $\theta$-greedy exploration.

Uncertainty quantification (UQ) in deep learning regression is of wide interest, as it supports critical applications including sequential decision making and risk-sensitive tasks. In heteroskedastic regression, where the uncertainty of the target depends on the input, a common approach is to train a neural network that parametrises the mean and the variance of the predictive distribution. Yet to this day, training deep heteroskedastic regression models poses severe practical challenges in the trade-off between uncertainty quantification and mean prediction, such as variance overfitting, optimization difficulties and representation collapse. In this work, we identify the core issues and propose a simple and efficient procedure that addresses these jointly by post-hoc fitting a variance model across the intermediate layers of a pretrained network on a hold-out dataset. We demonstrate that this method is competitive with end-to-end trained mean-variance networks in heteroskedastic UQ on several data modalities. The method retains mean prediction accuracy, is cheap at both train and prediction time, and requires no additional data compared to existing methods. Finally, we show this method works on large scale foundation models for chemistry, paving the way for cheap heteroskedastic UQ in large scale neural networks without having to retrain such models end-to-end.

Artificial Intelligence (AI) has been on the rise in many domains, including numerous safety-critical applications. However, for complex systems in the real world, defining the underlying environmental conditions in which the AI-based system must operate---the Operational Design Domain (ODD)---is extremely challenging. This often results in an incomplete description of the ODD, which contrasts with the requirements of many domains for certifying AI-based systems. Traditionally, the ODD is created in the early stages of the development process, drawing on sophisticated expert knowledge and related standards. This paper presents a novel Safety-by-Design method to a posteriori define the ODD from previously collected data using a multi-dimensional kernel-based representation. This approach is validated through both Monte Carlo methods and a real-world aviation use case for a future collision-avoidance system. Moreover, by defining under what conditions two ODDs are similar, the paper shows that the data-driven ODD can produce a dataset similar to the original, hidden ODD. Deriving the novel, Safety-by-Design, deterministic kernel-based affinity representation of ODDs is fully automated via a bounded, order-independent algorithm. Utilizing the proposed ODD representation enables future certification of data-driven, safety-critical AI-based systems.

Proteolysis-targeting chimeras (PROTACs) induce protein degradation through coordinated interactions among a degrader molecule, a target protein, and an E3 ubiquitin ligase. Yet computational prediction is often framed as if degradation were an intrinsic property of the degrader alone, overlooking the tuple-level context that determines activity. This mismatch is especially consequential in public PROTAC datasets, where degradation labels are sparse but molecule–target–E3 records are abundant. We introduce DegradeQuery, a tuple-conditioned framework that uses unlabeled records as structured relational evidence rather than incomplete labeled examples. DegradeQuery first performs counterfactual tuple pretraining, learning to distinguish observed molecule–target–E3 tuples from alternatives generated by replacing the target, the E3 ligase, or both. This objective induces a conditional compatibility representation before degradation labels are used, enabling the model to capture how degrader activity depends on biological context. The pretrained representation is then fine-tuned for high/low degradation prediction. On the official PROTAC-8K benchmark, DegradeQuery achieves 0.9065 AUROC and 0.8500 accuracy, outperforming reported semi-supervised results without pseudo-labeling, teacher models, distillation, or ensembles. Ablations, scaffold holdout, target–E3 holdout, and target-wise few-shot adaptation further support PROTAC degradation prediction as a tuple-conditioned rather than molecule-only problem.

Exemplar-based image editing applies a transformation defined by a source-target image pair to a new query image. Existing methods rely on a pair-of-pairs supervision paradigm, requiring two image pairs sharing the same edit semantics to learn the target transformation. This constraint makes training data difficult to curate at scale and limits generalization across diverse edit types. We propose Delta-Adapter, a method that learns transferable editing semantics under single-pair supervision, requiring no textual guidance. Rather than directly exposing the exemplar pair to the model, we leverage a pre-trained vision encoder to extract a semantic delta that encodes the visual transformation between the two images. This semantic delta is injected into a pre-trained image editing model via a Perceiver-based adapter. Since the target image is never directly visible to the model, it can serve as the prediction target, enabling single-pair supervision without requiring additional exemplar pairs. This formulation allows us to leverage existing large-scale editing datasets for training. To further promote faithful transformation transfer, we introduce a semantic delta consistency loss that aligns the semantic change of the generated output with the ground-truth semantic delta extracted from the exemplar pair. Extensive experiments demonstrate that Delta-Adapter consistently improves both editing accuracy and content consistency over four strong baselines on seen editing tasks, while also generalizing more effectively to unseen editing tasks. Code will be made publicly available.

Video multimodal large language models have made strong progress on open-ended video understanding, but they still lack precise local spatiotemporal perception. When two videos share almost the same global semantics and differ only in a short time span or a small region, current models often fail to find the change and provide reliable evidence. We propose DeltaVid, a verifiable proxy-task framework that enhances fine-grained spatiotemporal perception with cross-video differences. The key idea is to turn cross-video spot-the-difference into a trainable perception signal, where a model identifies local changes, judges temporal boundaries, and organizes spatial evidence by comparing similar videos. To make this signal scalable to train and reliable to evaluate, we further introduce DeltaVid-10K and DeltaVid-Bench, which convert controllable local differences in real videos into evidence-labeled training and test samples. Experiments show that DeltaVid substantially improves performance on cross-video difference understanding and transfers the learned local evidence ability to general video understanding benchmarks, including MMVU, MLVU, Video-MME, VideoHolmes, VideoMMMU, LVBench, TempCompass, and LongVideoBench. These results show that cross-video differences are not only an effective way to diagnose fine-grained perception failures, but also a scalable proxy supervision that moves Video MLLMs from coarse semantic understanding toward fine-grained spatiotemporal evidence reasoning.

Fine-scale-faithful neural simulation under fixed storage budgets remains challenging. Many existing methods reduce high-frequency error by improving architectures, training objectives, or rollout strategies. However, under budgeted coarsen-quantize-decode pipelines, fine detail can already be lost when the carried state is constructed. In the canonical periodic incompressible Navier--Stokes setting, we show that primitive and derived fields undergo systematically different retained-band distortions under the same operator. Motivated by this observation, we formulate Derived-Field Optimization (DerivOpt), a general state-design framework that chooses which physical fields are carried and how storage budget is allocated across them under a calibrated channel model. Across the full time-dependent forward subset of PDEBench, DerivOpt not only improves pooled mean rollout nRMSE, but also delivers a decisive advantage in fine-scale fidelity over a broad set of strong baselines. More importantly, the gains are already visible at input time, before rollout learning begins. This indicates that the carried state is often the dominant bottleneck under tight storage budgets. These results suggest a broader conclusion: in budgeted neural simulation, carried-state design should be treated as a first-class design axis alongside architecture, loss, and rollout strategy.

Reinforcement learning (RL) rewards are notoriously difficult to design and control, often leading to the model learning unintended behaviors such as eward hacking. One potential solution is to monitor for reward hacking and penalize it when detected; however, training against a monitor could lead to evasive behavior, and our general understanding of how to apply monitors effectively during training is limited. To study how best to use monitors to mitigate reward hacking, we introduce and open source three realistic environments where Qwen3-4B reward hacks: a coding environment hackable via test overwriting, a medical chat environment hackable via sycophancy, and a biography generation environment hackable via hallucination. We first focus on the coding environment, where we find that: (1) models can learn to evade highly accurate monitors by exploiting systemic flaws in probes and LLM judges; (2) monitors that leak more learning signal during RL suppress reward hacking but are more often evaded; and (3) including easier problems in training can decrease reward hacking. We apply our findings to build better reward hacking monitors for the medical chat and biography generation environments that improve upon naive baselines to reduce reward hacking rates across seeds from 70-100\% to 0\%. Our results demonstrate that our takeaways translate to new settings and that better monitor intervention designs are possible.

Graphic designs, such as posters, advertisements, and infographics, are an important medium for communicating information and shaping understanding. Unlike natural images, they consist of layered elements with explicit compositional order. However, existing object detection models treat these elements as an unordered set, leaving compositional order unexploited. To address this limitation, we present Detect Anything in Graphic Design (DAD), a model that formulates graphic design detection as compositional deconstruction. It decodes elements in compositional order, using lower-layer elements to better detect higher-layer ones. The key feature of DAD is amodal detection, which predicts the full bounding box of each element, including regions occluded by elements placed above it. Building on this formulation, we propose Element Relative Policy Optimization (EleRPO), which extends GRPO from sequence-level supervision to element-level optimization. EleRPO provides fine-grained training signals that capture how each detected element contributes to overall detection quality, and works synergistically with compositional order to improve detection performance. To support training and evaluation, we build a dataset of 10 million graphic designs. Experiments show that DAD outperforms all baselines and achieves human-level performance in amodal detection, supporting effective image-to-layer decomposition. EleRPO consistently improves over GRPO across nine detection benchmarks.

Vision Transformers (ViTs) have been widely adopted as visual encoders in multimodal models; however, the reliance on third-party pretrained checkpoints exposes these systems to backdoor attacks. Existing test-time detection methods rely on either the predicted label or intermediate embeddings, limiting their applicability across diverse tasks and incurring substantial computational overhead. In this work, we investigate how backdoor triggers affect the multi-head attention mechanism of ViTs and identify two distinct anomalous patterns: $\textit{complete attention hijacking}$, causing the attention to exhibit abnormally high inter-head spatial overlap, and $\textit{partial attention hijacking}$, causing an abnormally large variance in the concentration of the attention distribution. Motivated by these findings, we propose $\textbf{DetectViT}$, a test-time backdoor detection method that requires neither training data, model outputs, nor any prior knowledge of the trigger. DetectViT quantifies the two hijacking phenomena via an inter-head consistency score and an inter-head entropy variance score, with thresholds estimated from a small out-of-distribution (OOD) calibration set drawn from any source. Extensive experiments on four representative attacks, spanning diverse architectures (DeiT, CLIP, LLaVA) and downstream tasks (image classification and captioning), demonstrate that DetectViT substantially outperforms all baselines with as few as 64--128 OOD calibration samples, attaining a true positive rate of $100\%$ and a false positive rate of $0.17\%$ on BadCLIP. Notably, it introduces negligible additional overhead, thanks to its reliance solely on attention weights already computed during inference. We release our code at: https://anonymous.4open.science/r/DetectViT-81AC.

Aggregate math benchmark scores tell whether a model solved a problem, but not where reasoning broke down. We introduce milestone-oracle probing, a symbolically verified diagnostic that turns math evaluation into a reasoning-gap readout. For each parent problem, a teacher writes a fixed milestone roadmap. We evaluate a student under three levels of help: no help, the roadmap, and the roadmap plus gold milestone answers. We also run a separate milestone-only test, asking whether the student can solve each milestone in isolation. Deterministic symbolic verification grades all parent and milestone answers. Crossing these two measurements yields five gap categories: roadmap, milestone-execution, composition, missing-milestone, and capability. Applied to 354 NuminaMath problems and a six-model panel from 8B to 671B parameters, the largest residual category across all six models is families that pass the milestone test but are not recovered under any of the three probes (33--48\% of families per model). A follow-up audit on a broader sample of unrecovered families shows the residual is mixed: it contains real composition failures together with decomposition gaps and verifier artifacts. The protocol surfaces the residual; the audit reads it into its component parts. We release the diagnostic set, prompt templates, symbolic verifier, audit annotations, and the leak-safe repair logs and bookkeeping scripts.

Post-Training Quantization (PTQ) has become a prerequisite for efficient LLM deployment; however, current rotation-based methods are approaching a saturation point because they prioritize heuristic approximations over theoretical rigor. By treating quantization as an opaque black box, existing paradigms typically depend on indirect gradient estimators for activations while restricting weight optimization to local reconstruction objectives. Consequently, this structural limitation induces fundamental deficiencies: specifically, optimization instability and a "twofold blindness" toward accumulated upstream distortion and final task loss sensitivity. To overcome these barriers, we introduce $\textbf{DICEQuant}$, a unified framework established upon rigorous error modeling. First, we propose the $\textbf{CURE (Coupled Underflow-Rounding Error) surrogate}$, which facilitates exact and stable gradient computation for activation reshaping, thereby eliminating the variance inherent in heuristic estimators. Simultaneously, for weight optimization, we present $\textbf{Distortion-Compensated Rounding (DCR)}$. This mechanism derives a Shifted Optimal Center to neutralize input noise and utilizes Hessian-weighted shaping to align rounding decisions with the global functional objective. Empirical results demonstrate that DICEQuant significantly surpasses state-of-the-art baselines, successfully bridging the accuracy gap between low-bit compression and full-precision intelligence.

Human coordination often relies on the ability to influence the beliefs of others through strategic action. In multi-agent reinforcement learning, opponent shaping attempts to replicate this influence, though existing methods typically operate within an opponent's parameter, policy, or value space. Meanwhile, belief-manipulation techniques in hidden-role games often rely on hard-coded objectives, such as deception or belief saturation. We propose Differentiable Belief-based Opponent Shaping (D-BOS), a first-order method that treats each observer's belief as the shaped opponent state and differentiates through $k$-step softmax-Bayes belief dynamics. Rather than explicitly rewarding deceptive or cooperative behavior, our method treats the belief state as the target for shaping. This allows the optimal strategy to emerge naturally from the environment's reward structure. This belief-space formulation provides an opponent-shaping signal by differentiating through opponent belief updates, and naturally extends to multiple observers by aggregating gradients over their individual inferred belief trajectories. Empirically, D-BOS outperforms PPO and BBM in hidden-role games, with the largest gains in mixed-motive settings.

Score-based distillation methods train one-step diffusion models in two stages: they first train a teacher score model, then distill it into a one-step student model. However, this distillation process can introduce bias from two sources: errors in the teacher score model and student score estimate. We propose DiffRatio, a new framework for training one-step diffusion models without teacher supervision. Instead of using a teacher score model to provide training targets, DiffRatio directly learns a log density ratio between the student and data distributions across diffusion time steps. As a pre-trained score model is only used to initialize the one-step generator and does not supervise training. This design simplifies the training pipeline, mitigates gradient estimation bias, and reduces the size of auxiliary networks. In addition, the learned density ratio can be used as a verifier, enabling a principled inference-time parallel scaling method that further improves sample quality without external rewards or extra sequential computation. DiffRatio achieves strong one-step generation results on CIFAR-10 and ImageNet ($64{\times}64$ and $512{\times}512$), outperforming most teacher-supervised distillation methods.

We introduce a novel framework that transforms the resource-intensive (adversarial) prompt optimization problem into an efficient, amortized inference task. Our core insight is that pretrained, non-autoregressive generative LLMs, such as Diffusion LLMs, which model the joint distribution over prompt-response pairs, can serve as powerful surrogates for prompt search. This approach enables direct conditional generation of prompts, effectively replacing costly discrete optimization with a small number of parallelizable samples. We provide a probabilistic analysis demonstrating that under mild fidelity assumptions, only a few conditional samples are required to recover high-reward (harmful) prompts. Empirically, our method substantially outperforms prior attacks in both success rate and compute cost, producing low-perplexity, diverse jailbreaks that transfer to a wide range of black-box target models, including robustly trained and proprietary LLMs. Beyond adversarial prompting, our framework opens new directions for red teaming, automated prompt optimization, and leveraging emerging Flow- and Diffusion-based LLMs.

Residual connections with PreNorm are the default design in modern transformers due to their strong optimization stability, but recent research has shown that they weaken the functional differentiation of deeper blocks and increase representational redundancy in later layers. This issue also appears in diffusion models using similar architectures, where the deep block is expected to support progressive denoising refinement rather than repeated computation over similar states. Motivated by this, we revisit cross-layer shortcuts in diffusion transformers and argue that their role should go beyond merely transferring shallow information, instead guiding deeper blocks to improve depth utilization. To this end, we propose \textit{\textbf{Res}idual Adaptive Layer \textbf{Norm}alization} (\textbf{ResNorm}), which routes residual cross-layer shortcuts into the existing adaLN modulation. ResNorm preserves the optimization advantages of PreNorm residual connections, while reusing the native adaLN interface to transmit early-layer representations as modulation signals. This design naturally combines cross-layer history with the model's native condition, such as timestep and class embeddings. ResNorm allows the earlier structural representation to adaptively scale, shift, and gate deeper blocks. This not only produces more differentiated deep representations and reduces cross-layer similarity in later blocks, but also accelerates training convergence and improves performance. On ImageNet 256$\times$256, SiT-XL/2 + ResNorm achieves up to $\textbf{2}\times$ and $\textbf{36}\times$ faster training than SiT-XL/2 + REG and SiT-XL/2 + REPA. Moreover, SiT-L/2 + ResNorm trained for only 400K iterations already outperforms SiT-XL/2 + REG.

Reinforcement Learning with Verifiable Rewards (RLVR) has catalyzed significant advances in the reasoning capabilities of Large Language Models (LLMs). However, effectively managing the exploration and exploitation trade-off remains a critical challenge. In this paper, we fully analyze the exploration and exploitation dilemma of extremely hard and easy samples during the training and propose a new fine-grained trade-off mechanism. Concretely, we introduce a perplexity space disentangling strategy that divides the sample space into distinct exploration (high perplexity) and exploitation (low perplexity) subspaces, thereby mining fine-grained samples requiring exploration-exploitation trade-off. Subsequently, we propose a bidirectional reward allocation mechanism with a minimum impact on verification rewards to implement perplexity-guided exploration and exploitation, enabling more stable policy optimization. Finally, we have evaluated our method on two mainstream tasks: mathematical reasoning and function calling, and experimental results demonstrate the superiority of the proposed method, confirming its effectiveness in enhancing LLM performance by fine-grained exploration-exploitation trade-off.

Time series forecasting often suffers from over-smoothing, especially when future dynamics are multi-modal. Forecasts may follow the coarse trend of the observed future, but fail to preserve sharp changes, oscillations, turning points, and regime transitions that define plausible dynamic evolution. In this work, we revisit over-smoothing from the perspective of latent dynamical mode compression: under partial observation and single-realization supervision, multiple plausible future modes can be weakened, merged, or averaged during forecasting. Based on this view, we propose Dirichlet-Guided Group Forecasting (DGF), a mode-preserving forecasting framework that explicitly models multiple mode-conditioned predictive distributions and uncertainty over their selection probabilities. DGF uses a Dirichlet-guided hierarchical sampling mechanism and reward-based optimization to encourage forecasts that are accurate, dynamically consistent, and mode-distinct. Extensive experiments on real-world forecasting benchmarks show that DGF reduces over-smoothing while improving forecasting accuracy, diversity, and dynamical consistency.

While Gaussian Splatting-based Feature Fields (GSFFs) have shown promise for visual localization, this paper highlights that photometrically optimized GSFFs are inherently ill-suited for 2D-3D matching. The volumetric extent of each Gaussian induces many-to-one pixel-to-point mappings that destabilize PnP-based pose estimation, while photometric optimization gives rise to superfluous Gaussians devoid of multi-view consistency. To address these issues, we propose SplitGS-Loc, a localization-specialized GSFFs construction framework that disambiguates 2D-3D correspondences by exploiting Gaussian attributes. Our key design, Mixture-of-Gaussians-based splitting, decomposes each Gaussian into smaller Gaussians, replacing ambiguous many-to-one with precise one-to-one correspondences. In parallel, we exploit composition weights from GS rasterization to select Gaussians that significantly and consistently contribute across multiple views and aggregate discriminative features through strong pixel-Gaussian associations, enforcing multi-view consistency. The resulting compact yet discriminative feature fields enable stable PnP convergence, achieving state-of-the-art performance on localization benchmarks. Extensive experiments validate that SplitGS-Loc extends the utility of photometric GSFFs to accurate and efficient localization by exploiting Gaussian attributes, without per-scene training or iterative pose refinement.

Sequential Model-Based Optimization (SMBO) traditionally relies on Bayesian or ensembling surrogates for uncertainty quantification. While historically treated as fully data-driven, SMBO increasingly integrates external domain expertise to accelerate discovery. To overcome the opaque guidance and diminished integration fidelity of standard acquisition re-weighting, Probabilistic Circuits (PCs) have emerged as a generative surrogate alternative, enabling direct knowledge injection via conditional sampling. However, these generative routines lack the formal exploration-exploitation semantics required for rigorous optimization. We introduce the Distributional Conformance Score (DisCo), a principled metric that unifies the flexibility of PCs with a rigorous uncertainty framework. DisCo provides a bounded, $[0, 1]$-normalized measure of model "surprise" that (1) recovers properties comparable to kernel-based uncertainty known from, e.g., Gaussian Processes while maintaining linear-time inference, and (2) enables principled and accurate assessment of conformance of external knowledge w.r.t. model evidence. We then present DisCoMBO, a framework leveraging these properties for robust, knowledge-aware optimization. We prove that DisCoMBO is a zero-regret algorithm and demonstrate its effectiveness across diverse benchmarks from AutoML, material optimization, and wind park optimization.

When an RL agent's observations contain distractors driven by the same confounders as its true state, observational data alone cannot identify which dimensions the agent controls. In our benchmarks, even state-conditioned observational selectors can collapse when distractors mimic controllable state variables. We propose Interventional Boundary Discovery (IBD), which treats the agent's own action channel as a source of randomized interventions: randomizing actions implements an interventional contrast, and per-dimension two-sample tests with FDR correction produce a binary mask over observation dimensions. Across 12 continuous-control settings with up to 100 distractors, IBD matches oracle return in 11 of 12 settings, while observational baselines including mutual information, state-conditioned forward models, and gradient-based sensitivity often underperform simply passing the full observation to SAC.

Discrete diffusion is a promising framework for categorical generative modeling, but current evaluation protocols often fail to reveal whether models have actually learned the target distribution. We introduce the Discrete Diffusion Playground, a diagnostic benchmark that converts known 2D continuous distributions into binary token sequences through an exact discretized bijection. This creates a rare setting in which models are trained on discrete sequences, but generated samples can be decoded back into 2D and compared directly against a known ground-truth distribution. The benchmark spans smooth Gaussian mixtures and tilted checkerboards with sharp boundaries, disconnected support, and coordinate-dependent structure, allowing it to expose failures that are invisible in standard proxy metrics. It supports exact grid divergences, Wasserstein distances, density visualizations, and nonparametric two-sample tests, providing both quantitative evaluation and interpretable failure diagnosis. Using this testbed, we show that masked diffusion models can learn meaningful adaptive reveal orders but remain less stable than autoregressive baselines on smooth targets and largely fail on tilted checkerboards, where autoregressive models preserve the target geometry. Control experiments with reversed and interleaved token orderings show that autoregressive performance is not explained by favorable bit order alone.The Playground provides a simple but stringent benchmark for stress-testing discrete generative models and for developing evaluation metrics that detect real distributional mismatch rather than proxy performance. Our code is available at https://anonymous.4open.science/r/DiscreteDiffusionPlayground-0CCA/

The robust delivery of power is a non-negotiable prerequisite for the operational reliability of modern chips. Even a single unanticipated IR-drop violation can cause system-level crashes, making IR-drop a strict criterion in chip signoff. However, current methodologies ranging from physics-based simulators to recent learning-based surrogates remain bottlenecked by an inherent coupling between spatial topology and temporal evolution. This paradigm is intrinsically unscalable when billion-gate spatial complexity converges with the extensive workloads of full-testbench verification. In this work, we reformulate dynamic IR-drop assessment as a disentangled multimodal learning problem and introduce Layout Waveform Fusor (LWFusor), which decouples invariant spatially static chip layouts from transient switching waveforms and fuses them through physics-aligned end-to-end learning. By projecting activity sequences directly into the spatial potential domain, the proposed approach departs from traditional time-stepping bottlenecks and makes long-horizon power integrity analysis computationally efficient. Experiments on industrial-scale designs demonstrate orders-of-magnitude acceleration while maintaining high fidelity in hotspot capture, highlighting a new trajectory for power integrity sign-off in large-scale EDA workflows.

Foundation models for time-dependent partial differential equations (PDEs) are trained on large and diverse collections of physical systems and can generalize effectively to new downstream tasks. After fine-tuning on only a few trajectories from a target domain, they often achieve strong accuracy in low-data regimes. However, these models are typically large and computationally intensive, limiting their usefulness as fast surrogates for numerical solvers. We propose Teacher Rollout Extension (TREX), a knowledge distillation framework that transfers the predictive capability of a pretrained foundation model into a compact and efficient student. Starting from a fine-tuned teacher, TREX augments limited downstream data by generating long synthetic trajectories through teacher rollouts with periodic noise injection. This procedure samples from the teacher-induced rollout distribution without requiring explicit knowledge of the initial-condition distribution, while exposing the student to long-horizon states and local recovery behavior around states encountered during autoregressive prediction. The student can further incorporate task-specific inductive biases, such as equivariance, that the teacher does not necessarily enforce. We evaluate TREX on multiple PDE benchmarks. The resulting students match or even surpass the teacher’s accuracy while reducing the number of parameters by several orders of magnitude and achieving more than an order-of-magnitude speedup in inference.

GNN-to-MLP distillation aims to retain the predictive accuracy of a message-passing teacher while deploying a graph-free MLP at inference. Existing methods mainly transfer node-wise predictions or use confidence-based reweighting, but they do not specify where the student should preserve the teacher's graph-induced geometry. We show that this omission leads to two spectral failure modes in the student's representation space. On sparse graphs, the student suffers from *spectral underfit*, missing high-energy teacher directions concentrated near boundary regions. On dense graphs, it suffers from *spectral overfit*, retaining spurious directions that the teacher has collapsed through aggregation. Motivated by an energy-weighted teacher--student alignment objective, we propose **Graph Geometry-aware MLP (G$^2$MLP)**, a training-time distillation framework guided by Ollivier--Ricci curvature. Curvature identifies where the two spectral errors concentrate and is used to allocate supervision between prediction-level and representation-level alignment. The deployed model remains a standard MLP and requires no graph access at inference. Across node-classification benchmarks, G$^2$MLP consistently improves over graph-free distillation baselines, reduces the teacher--student rank gap in both regimes, and transfers without architectural changes to Graph Transformer teachers and link prediction.

Current end-to-end autonomous driving models are fundamentally constrained by the behavioral cloning ceiling of imitation learning. While reinforcement learning offers a path to smarter autonomy, it demands two missing pieces of infrastructure: (1) a cognitive foundation that understands traffic semantics and driving intent, and (2) a foresighted physical environment that can anticipate the consequences of candidate actions. To this end, we propose CoPhy, a Cognitive-Physical reinforcement learning framework for autonomous driving. To distill to think, we distill VLM knowledge into the BEV encoder and then discard the VLM entirely, retaining cognitive ability at zero inference cost while releasing the cognitive channel as a pluggable interface for optional human language commands. To foresee to act, we build an auto-regressive BEV world model that explicitly predicts future semantic maps conditioned on candidate actions, serving as an interpretable physical sandbox from which safety metrics are directly derived. Built upon this dual infrastructure, we optimize the driving policy via GRPO with a novel dual-reward mechanism: a physical reward derived from BEV rollouts enforces hard safety constraints, while a cognitive reward from a language-aligned scorer ensures intent compliance. Extensive experiments demonstrate that CoPhy not only achieves state-of-the-art results on NAVSIM v1 and v2 benchmarks, but also enables safer driving via cognitively informed scene compliance and flexible intent control through user-defined language instructions.

Most existing Spiking Neural Networks (SNNs) are based on first-order ordinary differential equations (ODEs) to model neuron potential dynamics. Recently, fractional-order spiking neural networks (f-SNNs) extend conventional SNNs by incorporating fractional-order ODEs, enabling the modeling of long-range temporal dependencies in neuron potential states. However, existing first-order SNNs and f-SNNs are typically restricted to a single derivative order, which limits their capacity to represent heterogeneous temporal dynamics across different time scales. In this work, we propose Distributed-Order Fractional Spiking Neural Networks (DF-SNNs), which generalize single-order neuron dynamics by integrating multiple fractional orders through learnable distributed weights. This formulation enables neurons to accumulate historical information across multiple temporal scales within a unified framework. We develop practical numerical implementations based on the Grünwald–Letnikov discretization, allowing efficient training of DF-SNNs within standard SNN architectures. We show that the proposed distributed-order formulation induces intrinsic multi-scale temporal memory, providing a richer class of temporal representations compared to single-order fractional dynamics. Extensive experiments demonstrate consistent performance gains of DF-SNNs over conventional SNNs and single-order f-SNNs.

In-context learning (ICL) allows large language models (LLMs) to adapt to new tasks through demonstrations, yet it suffers from escalating inference costs as context length increases. While task vectors offer a promising alternative by compressing demonstrations into compact hidden-state representations, their quality has been evaluated only through downstream task accuracy. This indirect criterion provides limited insight into how to design more effective task vector extraction methods. In this paper, we posit that inference using task vectors should align their predictive distribution with that of ICL. To quantify this, we introduce $d_{\text{NTP}}$, a metric that measures the discrepancy in next-token probabilities between task vector-based and ICL-based inference. Our empirical analysis reveals that $d_{\text{NTP}}$ serves as a performance proxy, exhibiting a strong negative correlation with downstream accuracy. Motivated by this, we develop Linear Task Vector (LTV), a method designed to minimize $d_{\text{NTP}}$ via a closed-form linear mapping that estimates demonstration effects through regression. Across eight classification benchmarks and five LLMs, LTV consistently outperforms existing task vector baselines, improving average accuracy by 9.2\% while reducing inference latency. We further show that LTV outperforms the baselines on regression tasks. Moreover, we investigate the transferability of LTV across different model scales; an aspect that has remained nascent in task vector research. Specifically, we empirically show that task vectors from a larger model can enhance a smaller model's performance by 6.4\%, suggesting a new utility for extracted task representations.

Mixture-of-Experts (MoE) transformers scale capacity by activating only a few experts per token, but this sparsity creates a hidden reliability problem: when routing is imperfect, load-balanced models may send tokens to experts that are insufficiently trained for the assigned inputs. We propose Distributionally Robust MoE Training (DRMoET), a drop-in objective that treats layer-wise experts as endogenous robustness groups and optimizes high-loss routing outcomes rather than merely equalizing traffic. DRMoET updates a per-layer expert distribution by an entropy-regularized softmax rule on EMA-smoothed, activation-weighted expert losses, strengthening plausible non-top routing paths while preserving standard MoE computation. Under the FLAME-MoE recipe at 746M-total and 10.3B-total scales, DRMoET improves downstream averages over both standard FLAME-MoE and auxiliary-loss-free balancing. Mechanistic analyses show lower expert-loss variance with nearly unchanged mean loss, 4.3\% lower degradation under forced mid-$k$ misrouting, and improved domain--expert specialization. These results position routing robustness--not only utilization balance--as a practical objective for reliable sparse MoE scaling.

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to autoregressive models, offering competitive performance while naturally supporting parallel decoding. However, as dLLMs are increasingly integrated with Mixture-of-Experts (MoE) architectures to scale model capacity, a fundamental mismatch arises between block parallel decoding and token-level expert selection. Specifically, each dLLM forward pass processes multiple tokens with bidirectional dependencies, whereas conventional MoE layers route each token independently. This mismatch substantially increases the number of uniquely activated experts, making inference increasingly memory-bound. To address this, we propose dMoE, a simple yet effective block-level MoE framework. The central idea of dMoE is to aggregate token-level expert distributions within each block into a unified block-level expert distribution, which is then used to guide expert routing in a more coherent manner. In this way, dMoE substantially reduces the number of uniquely activated experts during inference without sacrificing performance, thereby mitigating the memory-bound bottleneck. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of dMoE. On average, dMoE reduces the number of uniquely activated experts from 69.5 to 14.6 while retaining 99.11\% of the original performance. Meanwhile, it reduces memory usage by 76.64\% to 79.84\% and achieves 1.14$\times$ to 1.66$\times$ end-to-end latency speedup.

Vision-Language-Action (VLA) models are increasingly used as general-purpose policies for robotic manipulation, but completing the instructed task requires more than reaching the target state. A policy must also avoid nearby objects throughout the rollout, including approach, interaction, transport, and handover. A common approach is to invoke conventional collision avoidance mechanisms, such as motion planners or safety filters, which provide effective geometric safeguards. However, the resulting system may struggle to produce actions that are simultaneously safe and task progressing. When collision avoidance and manipulation progress require misaligned actions, the system may avoid the obstacle while slowing, disrupting, or even failing the task. To address this, we introduce DODGE-IT, a collision-aware learning framework that leverages collision signals as policy-improvement feedback to strengthen the VLA policy's collision-aware manipulation capability. DODGE-IT first designs Task-Relevant Assessment of Collision Events (TRACE) that detects and identifies task-relevant obstacles from RGB-D observations and turns executions into collision outcomes. These outcomes are then incorporated together with task success signals into representative reinforcement learning frameworks for VLA. We validate DODGE-IT on SafeLIBERO and real robot manipulation tasks, showing improved obstacle avoidance while preserving task completion across simulated and physical settings.

Diffusion-based visual generative models are powerful, but understanding their generalization behavior remains challenging. Current evaluations rely on human preference scores for text-to-image models and distribution-level metrics such as Frechet Inception Distance (FID) and Inception Score (IS) for class-conditional models, but these do not test whether models learn basic visual skills or largely match training patterns. To overcome this gap, we introduce the Visual Generative Lab (VGL), a controlled experimental framework for understanding how diffusion models learn and generalize basic visual skills, including size, position, rotation, count, color, shape, and their composition. By generating synthetic training data for each skill from explicit rules, VGL enables precise measurement of generalization using task-specific, rule-based metrics. Training models from scratch on these data reveals three consistent patterns. (1) Models extrapolate rotation better than size, skill, or count on far-out-of-range queries; healthy rotation seeds match most extrapolation angles within a few degrees, although a non-trivial fraction of seeds collapse near the wraparound, while skill and position saturate at small distances outside training regardless of seed. (2) For compositional generalization, coverage of basic visual skill combinations matters more than dataset size. (3) Basic visual skills are learned jointly, so an out-of-distribution request in one skill can degrade the others even when they remain in range.

Flow cytometry is the primary assay for single-cell immune profiling, yet cross-study modelling remains under-explored. The challenge is two-fold: Studies measure different marker panels, inducing feature-space inconsistency, while batch effects shift signals even on shared markers, making transfer harder than in vision or language where input spaces are stable. We conduct a large-scale empirical study, pretraining an encoder on 89,324 samples from 112 public studies spanning 525 distinct marker panels. We evaluate it frozen --- isolating representation quality from task-specific adaptation, on five clinical tasks under a taxonomy of panel-overlap conditions, including three panel-zero-shot tasks whose marker panels were unseen during pretraining, and one negative control with no expected discriminative signal. We further characterise transfer across scaling axes and probe robustness to marker removal at test time. Despite heterogeneity that disrupts direct transfer, the frozen encoder matches or exceeds in-distribution specialist encoders on 4 of 5 tasks, including all three novel-panel tasks, and behaves as expected on the negative control. Scaling analyses reveal saturation in both encoder capacity and random corpus subsampling, and representations degrade gracefully under marker reduction. These results suggest a pretrained encoder can serve as a general-purpose feature extractor for clinical cytometry data, without task-specific retraining.

AI systems increasingly assist decision making by producing cheap factor-level assessments of complex inputs, but these assessments are often biased and incomplete. We study selective human oversight: given AI signals, which factors should be escalated to a costly human evaluator? We formulate this human--AI collaboration problem as a factor-level information-acquisition problem. Under squared loss, solving for optimal dispatching reduces to maximizing a contextual reward; under a linear model, this reward admits a closed-form decomposition into two interpretable terms: predictive relevance and residual uncertainty in the human evaluation after conditioning on the AI signal. We instantiate the framework on peer review, decomposing papers into ten aspects and evaluating 3,408 ICLR submissions across three LLMs and multiple regression heads. Outputs based on the optimal dispatching rules match full-human-review performance using only 2--3 queried aspects out of 10. The framework remains effective when the human signal comes from a single noisy reviewer and across review years, suggesting robustness to reviewer inconsistency and and temporal distribution shift. These results position principled dispatching as a practical foundation for scalable human--AI decision systems.

Humans often say what sounds acceptable rather than what they truly think when facing social pressure, a well-documented phenomenon called \textbf{Social Desirability Bias (SDB)}. Trained on human-generated text, LLMs are steeped in the same social fabric that produces these pressures, yet whether and how SDB shapes their responses remains unknown. We present \textbf{PRESS} (\textbf{P}sychologically-grounded \textbf{R}epresentation \textbf{E}xtraction for \textbf{S}ocial-desirability \textbf{S}teering), the first framework to identify, localise, and steer SDB inside open-weight LLMs. We decompose SDB into four observable and relatively orthogonal mechanisms: \emph{Audience}, \emph{Accountability}, \emph{Self-Monitoring}, and \emph{Social Norms}, and operationalise each as paired High/Low pressure conditions across 9 cultures in cultural values where social pressure bites hardest. Our experiments show that SDB forms a decodable and low-dimensional direction inside LLMs, organised by specific psychological mechanism rather than cultural origin. Steering along it improves all 58 downstream cultural content moderation tasks across three model families (Qwen $+5.7\%$,Gemma $+7.9\%$, Llama $+16.3\%$ mean F1 across cultures), and a direction extracted from one culture can transfer to others without re-extraction. Ablation confirms the direction is causally necessary, and evaluation on capability benchmarks further shows PRESS shifts cultural stance without disrupting factual knowledge.We hope this work makes social pressure in LLMs more transparent and opens new directions toward more honest and socially aware language models.

Modern visual recognition systems increasingly rely on unlabeled images collected from diverse visual domains. A common goal in dataset construction is to avoid obvious marginal imbalance, such as skewed class totals or uneven domain sizes. However, marginal balance can be misleading in multi-domain data: even when per-class totals and per-domain sample sizes are balanced, sample allocations can be highly skewed across domain-class pairs. We call this hidden conditional skew Domain-Conditioned Class Imbalance (DCI), where class imbalance is revealed only after conditioning on domain context. For controlled evaluation, we fix class and domain marginals and vary only the joint domain--class allocation, making the effect of domain-conditioned imbalance directly measurable. To address this challenge, we propose HyperIDAC, a hypernetwork-based classifier for multi-domain imbalance. HyperIDAC infers domain priors from visual features and lightweight zero-shot predictions, then dynamically generates instance-specific classifier parameters and domain-conditioned class embeddings. This design yields domain-sensitive decision boundaries while leveraging CLIP semantics to support underrepresented domain--class pairs. Furthermore, a multi-level pseudo-labeling module combines teacher–student training with historical label tracking and a fallback for low-confidence samples, mitigating bias reinforcement in unlabeled settings. Extensive experiments on four multi-domain benchmarks under three regimes show that HyperIDAC substantially outperforms state-of-the-art methods, highlighting the value of domain-aware hypernetwork-based adaptation under DCI.

Activation steering controls LLM behaviour towards target behaviour by intervening in internal representations, yet it often degrades reasoning and retrieval performance. We argue that a primary cause of this trade-off is attention rerouting: steering vectors alter query-key matching, shifting attention away from contextually important tokens toward less informative ones. To address this, we propose Steering via Key-Orthogonal Projections (SKOP), a steering method that constrains harmful attention rerouting without eliminating steering efficacy. SKOP achieves this by preserving attention patterns on a small set of focus tokens the model relies on for reasoning and retrieval, while allowing redistribution among less critical tail tokens. Across multiple steering benchmarks, we show that SKOP achieves the best joint steering-utility trade-off, reducing utility degradation by 5–7$\times$ while retaining over 95\% of vanilla steering efficacy. Our results further suggest that, in long-context retrieval settings where vanilla steering approaches are ineffective, SKOP can maintain robust performance by avoiding attention rerouting.

Selecting the optimal LLM inference configuration requires evaluation across hardware, serving engines, attention backends, and model architectures, since no single choice performs best across all workloads. Profile-based simulators are the standard tool, yet they hardcode their operation set to a specific configuration and re-profile every operation from scratch, making exploration prohibitively expensive. This cost stems from a missing structural understanding: every input dimension of each operation is fixed by the model configuration or determined by the incoming request. Many model-configuration values (e.g., head size, layer count) recur across models, so the same operation runs in many configurations; a single sweep over the request-dependent dimensions can serve them all. We present Dooly, which exploits this structure to achieve configuration-agnostic, redundancy-aware profiling. Dooly performs a single inference pass, labels each input dimension with its origin via taint propagation, and selectively profiles only operations absent from its latency database; stateful operations such as attention are isolated by reusing the serving engine's own initialization code, eliminating manual instrumentation. It builds latency regression models based on the database, which becomes a drop-in backend for existing simulators. Across two GPU platforms, three attention backends, and diverse model architectures, Dooly achieves simulation accuracy within 5% MAPE for TTFT and 8% for TPOT while reducing profiling GPU-hours by 56.4% across 12 models compared to the existing profiling approach.

Modern distributed optimization methods mostly rely on traditional synchronous approaches, despite substantial recent progress in asynchronous optimization. We revisit Synchronous SGD and its robust variant, called $m$-Synchronous SGD, which can be interpreted as a method with a backup-worker strategy, and theoretically show that they are nearly optimal in many heterogeneous computation and communication scenarios, which is somewhat unexpected. We analyze the synchronous methods under random times and adversarial partial participation of workers, and prove that their time complexities are optimal in many practical regimes, up to logarithmic factors. While synchronous methods are not universal solutions and there exist tasks where asynchronous methods may be necessary, we show that they are sufficient for many modern heterogeneous computation and communication scenarios.

Generative object detection, introduced by DiffusionDet, is defined by two coupled properties: (P1) data-independent random-box initialization with a learned noise-to-data transport, and (P2) test-time scaling (TTS), the ability to trade inference compute for accuracy at a fixed checkpoint by adjusting the number of proposals or refinement steps. This paper asks: \emph{How can we preserve both properties with a minimal conceptual design?} We dissect DiffusionDet component by component and find that, in our controlled interventions, the diffusion schedule, the initial noise distribution, and DDIM sampling stochasticity have only a small effect on inference results. In contrast, detection performance is more sensitive to the denoising head, iterative refinement, and proposal coverage. Based on these diagnostic results, we propose \textbf{LinearDet}, a compact generative detection prototype: \emph{linear noising at training, linear denoising at inference}---no diffusion-specific schedule, no diffusion ODE/SDE solver, only a symmetric pair of linear mixtures. On COCO, LinearDet matches DiffusionDet under standard settings, is more robust under sparse proposal coverage, outperforms it in zero-shot CrowdHuman transfer, and produces smoother refinement trajectories that admit clean heatmap visualization. More broadly, we hope this diagnosis and compact prototype offer the object detection community a cleaner baseline for studying generative detection.

Modern image editors excel at semantic manipulation and visual synthesis, yet remain limited in precise spatial control, motivating the development of drag-based editing. However, existing drag-based methods often struggle to balance drag accuracy with natural, plausible, and intent-aligned generation. We propose MoRe-Drag, a motion-grounded drag-based editing method. Our key insight is to treat pixel-space warping as coarse motion evidence, and to inject this evidence into the generative sampling trajectory. Specifically, MoRe-Drag performs region-aware latent recomposition over refinement, inpainting, and anchor regions, coupled with stage-adaptive conditioning that progressively shifts from motion-grounded structure formation to semantic refinement. We further support an instruction-free interface by adapting the MLLM-based text encoder for drag-aware instruction inference. Experiments on DragBench-SR and DragBench-DR show that MoRe-Drag substantially improves drag precision over strong base editors and achieves superior drag accuracy among SOTA drag-based methods, while delivering strong semantic consistency and visually realistic results. Code and dataset will be publicly released.

Diagram question answering (DQA) requires models to interpret structured visual representations such as charts, maps, infographics, circuit schematics, and scientific diagrams. Recent vision-language models (VLMs) often achieve high answer accuracy on these tasks, yet correct answers do not guarantee that models ground their reasoning in the diagram regions that support the prediction. Models may instead rely on textual correlations or dataset artifacts without identifying the visual evidence required to verify the answer. This limitation prevents reliable evaluation of diagram reasoning and reduces interpretability. We introduce DRAGON, a benchmark for evaluating evidence-grounded visual reasoning in diagrams. Given a diagram, a question, and the correct answer, a model must predict bounding boxes that correspond to the visual elements required to justify the answer. These evidence regions may include answer-bearing components, textual labels, legends, axes, connectors, and other supporting structures involved in the reasoning process. The DRAGON dataset contains 11,664 annotated question instances from six diagram QA datasets: ChartQA, Circuit-VQA, InfographicsVQA, MapIQ, MapWise, and AI2D, with a 2,445-instance test set carrying human-verified evidence annotations and a standardized evaluation framework. Our evaluation on recent VLMs reveals that even Claude Opus 4.6 achieves a Grounding IoU of only 23.8% on AI2D, showing that faithful visual grounding remains an open challenge. DRAGON supports future research on models that ground their predictions in visual evidence.

We introduce Dream-Cubed, a large-scale dataset of Minecraft worlds at voxel resolution, and a family of models using cubes as powerful compositional units for efficient generation of interactive 3D environments. Dream-Cubed comprises tens of billions of tokens from a carefully curated mixture of procedural biome terrain and high-quality human-authored maps. We use this dataset to conduct the first large-scale study of 3D diffusion models for voxel generation, analyzing discrete and continuous diffusion formulations, data compositions, and architectural design choices. Our models operate directly in the space of blocks, enabling efficient and semantically grounded generation while supporting interactive user workflows such as inpainting and outpainting from user-authored blocks. To quantitatively evaluate our models, we adapt the FID metric to assess semantic differences between real and generated world renderings, and validate generation quality through a human preference study. We release the full dataset, code, and all our pretrained models, which we hope will provide a foundation for future research in efficient generative modeling for structured, interactive 3D environments.

We propose Drifting Field Policy (DFP), a non-ODE one-step generative policy built on the drifting model paradigm. Under online RL finetuning, we frame the policy update as a reverse-KL Wasserstein-2 gradient flow toward a soft target policy, realized by DFP as a steepest-descent step on probability space. By construction, this gradient step is decomposed into an ascent toward higher action-value regions and a score matching with the anchor policy as a trust region. We further derive a simple, tractable surrogate of the otherwise intractable update loss, akin to behavior cloning on top-K critic-selected actions. We find empirically that this mechanism uniquely benefits the drifting backbone owing to its non-ODE parameterization. With one-step inference, DFP achieves state-of-the-art performance on several manipulation tasks across Robomimic and OGBench, outperforming ODE-based policies.

Offline reinforcement learning requires improving a policy from fixed data while avoiding out-of-distribution actions with unreliable value estimates. Diffusion and flow policies handle this trade-off by modeling the behavior distribution to regularize the RL objective, but they require iterative denoising, solver integrations, and in more efficient variants, distillation or other approximations at inference. We propose DriftQL, which combines a drift-based behavioral regularizer with critic-driven policy improvement. The value signal biases the policy toward high-value regions of the data support, while attraction and repulsion together keep generated actions near the data and prevent collapse onto a single mode. DriftQL is implemented as a single network with a unified training objective and generates actions in a single forward pass. On D4RL and OGBench, DriftQL consistently outperforms diffusion and flow methods, advancing the state of the art. Under degraded data quality, where the baselines visibly struggle, DriftQL remains close to its clean-data performance, positioning it as a promising alternative to diffusion and flow-based methods while maintaining the simplicity and efficiency of deterministic approaches.

Single-stage fine-tuning is pervasive, yet catastrophic forgetting persists when the source pre-training data is inaccessible. Data weighting affords a natural lever to modulate sample influence. Yet existing weighting methods lack a principled mechanism to explicitly arbitrate knowledge retention against new-task adaptation, and provide no unified differentiable objective to accommodate auxiliary signals such as sample diversity and sample characteristics. This is because current schemes privilege one goal over the other, and no explicit gradient-compatible encoding exists for such signals. To address this, we propose $\textbf{DRIVE}$, a principled and efficient weight assignment framework. It derives per-sample weights via a unified differentiable objective where three coupled signals co-evolve: Shapley additivity splits contribution into anti-forgetting and adaptation components, which are reshaped by a covariance-based diversity term and stabilized by a sample-characteristic prior. Far from a mere combination, this yields an end-to-end gradient-based scheme balancing stability against plasticity. Theoretically, we prove that DRIVE enjoys superior stability-plasticity guarantees over existing weighting schemes and enhances parameter estimation robustness. Empirically, it attains superior overall performance, faster convergence, and stronger robustness. Code at https://anonymous.4open.science/r/DRIVE-158.

Existing latent world models for autonomous driving have opened a promising path toward future-aware driving intelligence. However, they typically treat future latent states as prediction targets or auxiliary signals, rather than directly conditioning trajectory planning. This can entangle current and future features in latent space. In this work, we propose DriveFuture, a future-aware latent world modeling framework for autonomous driving that explicitly learns planning-oriented foresight by conditioning the current latent state modeling process on future world states. Specifically, during training, the model first predicts future latent world states from the current latent state and ego action, and then refines the prediction against the ground-truth future latent state via cross-attention. The resulting future-aware latent serves as an explicit condition for a diffusion-based trajectory planner. During inference, DriveFuture conditions on the predicted future latent state instead of the ground-truth future state. DriveFuture achieves SOTA performance on the public NAVSIM benchmarks, reaching 55.5 EPDMS on NAVSIM-v2 navhard, 89.9 EPDMS on NAVSIM-v2 navtest, and 90.7 PDMS on NAVSIM-v1 navtest, respectively. These results suggest that the key to latent world modeling lies not merely in simulating future states, but more importantly in conditioning current decision-making on future states. Notably, as of April 2026, DriveFuture ranks 1st on the NAVSIM-v2 navhard leaderboard and achieves SOTA performance on NAVSIM-v1 navtest.

Multi-agent LLM systems improve reasoning and tool use via role specialization, yet reinforcement learning (RL) post-training for such systems remains unstable and underexplored. We theoretically pinpoint a key source of instability when extending group-based RL to cooperative multi-agent LLM systems: under GRPO-style optimization, a global normalization baseline can mismatch heterogeneous agents' reward distributions, inducing gradient-norm instability. Based on this finding, we propose Dr. MAS, a simple and stable RL recipe for cooperative multi-agent LLM systems. Algorithmically, Dr. MAS normalizes advantages per agent using each agent's own reward statistics, which calibrates gradient scales and dramatically stabilizes the training. Systemically, Dr. MAS provides an end-to-end multi-agent RL framework with scalable orchestration, flexible per-agent LLM serving and optimization, and shared resource scheduling of actor backends. Unlike single-actor frameworks such as veRL, Dr. MAS natively supports multiple heterogeneous LLMs over a unified GPU pool. Lifecycle-aware backend management and dynamic dispatch release inactive models' GPU memory and schedule active agents on demand, enabling hardware-efficient co-training. We evaluate Dr. MAS on multi-agent math reasoning and multi-turn search with Qwen2.5 and Qwen3. Dr. MAS achieves clear gains over vanilla GRPO (e.g., +5.6\% avg@16 and +4.6\% pass@16 on math, and +15.2\% avg@16 and +13.1\% pass@16 on search) while largely eliminating gradient spikes. It also remains highly effective under heterogeneous agent-model assignments while improving efficiency. Code is available at https://anonymous.4open.science/r/DrMAS-0787.

Building state-of-the-art (SOTA) predictive models for drug discovery requires expensive search over tools, architectures, and training strategies. Current LLM-based agents can find SOTA solutions through extensive trial and error, but they do not retain the experience accumulated along the way and therefore pay the full search cost on every new task. We propose DrugSAGE (Self-evolving Agent Experience), a framework that accumulates and reuses experience across tasks to build SOTA drug discovery models efficiently. DrugSAGE maintains a cross-task memory of verified skills, statistical evidence about effective strategies, and a record of recurring errors and their fixes. In some cases, DrugSAGE transfers a working solution directly without test-time search. In 33 molecular property prediction tasks, DrugSAGE ranks first among nine SOTA agents in a single-task setting. With memory accumulated from 16 smaller tasks, DrugSAGE achieves a averaged normalized score of 0.935 on 17 held-out tasks in a cross-task evaluation setting and outperforms all baseline agents by 10-30\% in a zero-test-time search regime. In summary, our work shows the advantage of cross-task memory for efficient SOTA model development in drug discovery.

Multi-agent trajectory prediction is crucial for ensuring autonomous driving safety. Although diffusion models show great potential in trajectory refinement, existing diffusion-based refiners suffer from severe exposure bias. This bias stems from the distribution mismatch between the noisy training endpoint and the noise-free inference starting point, which frequently causes negative optimization during the refinement process. To resolve this issue, this paper proposes the Diffusion Schrödinger Bridge Trajectory Refiner (DSBTR). By introducing a tractable Schrödinger bridge with dual-terminal constraints, DSBTR ensures that the forward evolution endpoint strictly degenerates into a deterministic and noise-free coarse prediction prior. This theoretical alignment eliminates the initial boundary error and constrains the reverse solving process to a strict residual approximation. Based on the semi-linear structure of DSBTR, we derive a tailored high-order DPM-Solver to accelerate the sampling process and meet autonomous driving application requirements. Extensive experiments on the ETH/UCY and SDD benchmarks demonstrate that DSBTR serves as an effective plug-and-play module. It consistently improves the prediction accuracy of various baseline architectures and successfully avoids negative optimization.

The rapid proliferation of personalization techniques in Stable Diffusion has driven the widespread deployment of adversarial perturbations as a protective measure against unauthorized data mimicry. Existing diffusion-based purification methods primarily rely on the noising and denoising paradigm and have proven somewhat effective in neutralizing these protections. However, they suffer from a severe trade-off between purification efficacy and perceptual fidelity, especially under large perturbation budgets. To mitigate this problem, in this paper, we propose Dynamic Sparse Semantic Anchoring (DSSA), a novel purification framework. By starting the reverse denoising process from pure Gaussian noise, DSSA removes the initial adversarial perturbations and circumvents the dilemma of selecting an optimal noising level. To prevent the loss of original semantics caused by pure noise generation, we propose a dynamic sparse anchoring strategy that directly integrates sparse pixels from the adversarial sample to guide the reconstruction. Additionally, we design a perception-guided early-stopping mechanism to prevent the perturbation resurgence caused by this direct pixel integration. Extensive experiments across diverse datasets and protection schemes demonstrate that our method consistently outperforms existing baselines, achieving state-of-the-art performance in efficacy and fidelity, successfully neutralizing even advanced adaptive protections.

Dynamic scene representation and rendering in the visible spectrum has been extensively studied. Compared to visible-light imaging, thermal infrared sensing offers all-weather observation and strong penetration capability, enabling perception in low-visibility environments. However, dynamic reconstruction in thermal infrared scenes is challenged due to atmospheric transport effects and interference from low-emissivity background regions, which often leads to motion blur and details loss in rendered results. To address these issues, we propose DTGS, a physics-driven dynamic 3D scene reconstruction approach. DTGS factorizes dynamic thermal scenes into radiometric and geometric components, modeling radiometric changes via rendering that integrates atmospheric radiative transfer, while capturing geometric deformations through Adaptive Temporal Gaussian basis. In addition, we design a thermal-radiance-weighted structural similarity loss to suppress gradient interference from low-radiance noisy regions. Furthermore, to demonstrate the effectiveness of our method, a large-scale dataset for this field named Dynamic_LTR is created. Experimental results demonstrate that our method achieves a 2.7 dB improvement in PSNR, a twofold speedup in rendering, and a threefold reduction in training time. The code and datasets will be available after acceptance.

Existing graph neural network-based methods for Boolean satisfiability (SAT) solving are generally constrained by limited local receptive fields, making it difficult to adequately capture the long-range dependencies between variables and clauses in SAT instances. Moreover, these methods overlook the multi-solution nature of SAT instances. Consequently, their expressive power and generalization performance remain limited. To address these issues, we propose DualSAT, a novel neural framework for the SAT problem, designed to enhance the phase selection heuristic of SAT solvers. Specifically, DualSAT adopts a novel dual-branch graph neural network that separately encodes local structural information and global structural information in the graph, and then fuses them through an attention-based cross-branch feature fusion module. An assignment decision head is further introduced to generate the initial assignment prediction of variables. This prediction process requires only a single forward pass, achieving a favorable balance between expressive power and inference efficiency. Furthermore, to account for the multiple solutions nature of SAT instances, we design a two-stage training strategy that combines supervised and unsupervised learning to improve the generalization performance of the model. Experimental results on multiple datasets show that, as an end-to-end assignment prediction model, DualSAT significantly outperforms NeuroSAT and NeuroBack. In addition, we integrate DualSAT into the classical SAT solver CaDiCaL, where it increases the number of solved instances by up to 11.6% and reduces the number of conflicts by up to 9.21% on SAT competition problem sets.

This paper develops a dual-space preconditioning framework for variational inequalities and root-finding problems beyond classical cocoercivity and Lipschitz continuity assumptions. To this end, we introduce both a novel cocoercivity-type condition and a relaxed Lipschitz continuity condition, and derive convergence of deterministic and stochastic methods under these generalized conditions. For isotropic preconditioners, the resulting algorithms can be interpreted as variable-stepsize variants of classical methods, recovering and extending recent analyses based on $(L_0, L_1)$-type growth conditions while allowing simpler proofs and larger stepsize ranges. We also show that the framework naturally handles constrained variational inequalities through weak Minty-type arguments. Finally, we study the associated continuous-time dynamics and derive a variational and optimal-control interpretation of the preconditioned flow. Overall, the results show that dual-space preconditioning provides a flexible mechanism for designing and analyzing operator methods beyond the standard Euclidean Lipschitz setting.

Class-Incremental Learning (CIL) requires models to continuously acquire new classes without forgetting previously learned ones. A dominant paradigm involves freezing a pre-trained model and training lightweight, task-specific adapters. However, maintaining task-specific parameters hinders knowledge transfer and incurs high retrieval costs, while naive parameter fusion often leads to destructive interference and catastrophic forgetting. To address these challenges, we propose Dynamical Adapter Fusion (DAF) to construct a single robust global adapter. Grounded in the PAC-Bayes theorem, we derive a fusion mechanism that explicitly integrates three components: the optimized task-specific adapter parameters, the previous global adapter parameters, and the initialization parameters. We utilize the Taylor expansion of the loss function to derive the optimal fusion coefficients, dynamically achieving the best balance between stability and plasticity. Furthermore, we propose a Robust Initialization strategy to effectively capture global knowledge patterns. Experiments on multiple CIL benchmarks demonstrate that DAF achieves state-of-the-art (SOTA) performance.

Mobile and wearable sensing offer a scalable pathway to longitudinal mental health monitoring, yet accurate prediction remains limited by two fundamental forms of distribution shift in human behavior. The first arises across users, where similar sensor patterns can carry different clinical meanings, a phenomenon we call inter-subject heterogeneity. The second arises within a single user, where behavioral baselines themselves evolve over time, producing intra-subject behavioral drift. In this domain, ground-truth labels come from Ecological Momentary Assessment (EMA), brief self-reports collected in everyday life that are intrusive and inherently sparse. As a result, approaches that address these distribution shifts by updating model parameters are difficult to apply. To address this gap, we propose DyCon (Dynamic Context Modeling), a dual-loop framework that personalizes by building and evolving a separate natural-language personal context for each user, without updating any model parameters. Within each session, a momentary refinement loop restructures this personal context to resolve inter-subject heterogeneity. Across sessions, a longitudinal update loop integrates only validated insights to track intra-subject behavioral drift. Extensive experiments on four public benchmarks (GLOBEM, PMData, StudentLife, LifeSnaps) show that DyCon consistently outperforms baseline methods, particularly under sparse supervision and behavioral drift. Our results suggest that evolving an explicit personal context offers a practical pathway for longitudinal personalization in real-world health monitoring.

Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, yet real-world deployment often requires continual capability expansion across sequential tasks. In such scenarios, Multimodal Continual Instruction Tuning (MCIT) aims to acquire new capabilities while limiting catastrophic forgetting. Existing methods mainly follow a module-composition paradigm: they maintain task-level prompts or LoRA experts and dynamically route or aggregate a subset of them at inference. However, samples within the same task can still differ substantially in visual scenes, question intents, and reasoning demands. This motivates instance-level adaptation to individual query-image pairs rather than only selecting or combining task-level modules. To this end, we propose DRAPE (Dynamic Cross-Modal Prompt Generation), a prompt-learning framework that synthesizes continuous instance-specific soft prompts for MCIT. Instead of selecting prompts from a fixed pool, DRAPE derives prompt queries from the textual instruction and cross-attends to visual patch features, producing query-image conditioned prompts that are prepended to the frozen LLM. To mitigate forgetting during sequential updates, DRAPE applies null-space gradient projection to the shared projector and uses CLIP-based prototype routing for task-label-free generator selection at inference. Extensive experiments on MCIT benchmarks show that DRAPE achieves state-of-the-art performance among representative prompt-based and LoRA-based continual-learning baselines.

Large-scale foundation models achieve strong performance across diverse tasks, but their size makes inference costly, largely due to dense matrix multiplications. Prior work reduces this cost by replacing dense weight matrices with efficient structured forms such as low-rank factorizations. However, these methods approximate weights rather than the output activations that determine inference accuracy. Consequently, small weight-space errors can be amplified by input activations, producing large output errors. In this work, we propose DyRA, an input-adaptive method that improves structured matrix multiplication approximation by correcting residual output errors during inference. We show that matrix multiplication can be approximated more effectively by directly optimizing low-rank factors of the output. DyRA builds on this insight by dynamically approximating and correcting the output error introduced by structured weight approximations. This combines efficient structured computation with input-dependent correction, yielding a more faithful approximation of full matrix multiplication under the same computational budget. Across vision, speech, and language models, DyRA consistently improves the accuracy–efficiency trade-off over structured weight approximations alone. Notably, DyRA achieves a $1.5\times$ end-to-end GPU speedup for DINOv3 while reducing accuracy degradation by more than $3\times$ relative to weight-only baselines.

Memory capacity is a critical factor determining the performance of Vision-Language-Action (VLA) models in long-horizon manipulation tasks. Existing memory-augmented architectures primarily rely on linear or flat storage, lacking structural priors for manipulation categories and hierarchical organization. This deficiency hinders efficient experience retrieval and limits generalization to unseen long-horizon task compositions. Inspired by the hierarchical organization of human experience, we propose ECHO (Experience Consolidation and Hierarchical Organization), a novel memory framework operating within a Continuous Hierarchical Space. By employing a hyperbolic autoencoder, ECHO maps VLA hidden states into this space. Leveraging hyperbolic metrics and entailment constraint mechanisms, experience vectors are organized into a semantic memory tree that supports efficient top-down retrieval. In parallel, a background consolidation mechanism continuously refines the memory tree through geometric interpolation and structural splitting, supporting virtual memory synthesis in the continuous space. We integrate ECHO into the $\pi_0$ foundation model. Evaluations on LIBERO and preliminary real-world experiments demonstrate the effectiveness of our approach, notably achieving a 12.8% absolute improvement in execution success rate over the $\pi_0$ baseline on LIBERO-Long, while improving compositional generalization on cross-suite unseen long-horizon tasks.

Recent advances in image editing have enabled fine-grained and highly controllable visual modifications.However, achieving similar levels of controllability in video editing remains an open challenge, particularly for precise local edits, dynamic temporal effects, and consistent lighting across frames.In this paper, we propose EditDistill, a framework that distills image-level visual transformations from edits into a compact descriptor to guide video editing.Specifically, given a source video and an edit prompt, we first generate an edited version of the initial video frame using an off-the-shelf image editing model. We then distill the visual transformation of this original-edited image pair into a compact representation termed an edit descriptor.This descriptor is injected into a video diffusion model through our lightweight fusion module, which steers the denoising process toward the desired edit. To align the video trajectory with the demonstrated image edit while preserving motion and layout, we train with edit-descriptor alignment and DINO structural consistency losses, and derive an equivalent inference-time guidance formulation under flow matching.Extensive experiments demonstrate that our method achieves strong performance on standard benchmarks and produces high-quality video editing results across diverse scenarios.

Training Neural ODEs requires backpropagating through an ODE solve. The state-of-the-art backpropagation method is recursive checkpointing that balances recomputation with memory cost. Here, we introduce a class of algebraically reversible ODE solvers that significantly improve upon both the time and memory cost of recursive checkpointing. The reversible solvers presented calculate exact gradients, are high-order and numerically stable -- strictly improving on previous reversible architectures. On scientific modeling and time series classification experiments, reversible solvers reduce the training time by more than $2\times$ while using at least $10\times$ less memory on average than recursive checkpointing.

We study the classical problem of linear regression under heteroskedastic noise, where observations arise from multiple sources with unknown and potentially unequal variances. Specifically, we consider a setting with $n$ sources, each providing a linear measurement of an unknown $d$-dimensional parameter vector $\beta^*$, where each source is associated with an unknown noise variance. We focus on the classic "subset-of-signals'' model, where one assumes that $m$ of the sources have a relatively low noise, formally, with variance of at most 1. Our goal is to accurately estimate $\beta^\*$ despite the presence of high-variance sources. We show that if $m$ is sufficiently large, it is possible to recover $\beta^*$ with sub-constant error. Specifically, we prove a recovery error bound of $\tilde{O} \left(\frac{\sqrt{nd}}{m}\right)$, significantly improving upon previously known results for this setting.

The computational complexity of anonymous games has been a central theme in algorithmic game theory, with foundational contributions by Daskalakis and Papadimitriou [2015], Goldberg and Turchetta [2017], and Cheng et al. [2017] establishing that the landscape of equilibrium computation admits a PTAS. In this paper, we investigate the class of $n$-player anonymous games with linear utilities and two actions. While finding an equilibrium in anonymous games is PPAD-hard for arbitrary utility functions—even with a constant number of actions [Chen et al., 2015]—we show that this linear structure allows for significant algorithmic improvements. Exploiting the fact that payoffs depend solely on the first moment of the players’ strategy distribution, we provide the first algorithm for computing exact Nash equilibria that runs in time polynomial in the number of players $n$. We achieve this by establishing a novel combinatorial characterization of the Nash equilibria. Finally, we establish a novel connection between Nash equilibria in these games and non-convex non-concave min-max optimization. Despite the technical challenges posed by this structural property, we design a sequential best-response dynamic that provably converges to an $\epsilon$-Nash equilibrium in $\mathcal{O}(\frac{1}{\epsilon})$ steps.

Assessing the capabilities and risks of frontier AI systems is a critical area of research; recent work has shown that repeated sampling can dramatically increase both. For instance, repeated sampling allows models to solve difficult math and coding problems, but it may also increase their risk of being jailbroken. Such results raise a crucial question: how can one accurately predict a model's behavior when scaled to a massive number of attempts, given a vastly smaller sampling budget? This question is directly relevant to model providers, who serve hundreds of millions of users daily, and to governmental regulators, who seek to prevent harms. To answer this question, we make three contributions. First, we find that standard methods for fitting these laws suffer from statistical shortcomings that hinder predictions, especially in data-limited scenarios. Second, we remedy these shortcomings by introducing a robust estimation framework, which uses a beta-binomial distribution to generate more accurate predictions from limited data. Third, we propose a dynamic sampling strategy that allocates a greater budget to harder problems. Combined, these innovations enable more reliable prediction of rare risks and capabilities at a fraction of the computational cost.

Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they still suffer from overthinking, generating redundant reasoning steps which incur substantial token consumption. Existing methods, such as suppressing reflective keywords or forcing shorter reasoning lengths, attempt to mitigate this issue but inevitably truncate necessary steps and induce underthinking, thereby compromising performance. To address this dilemma, we investigate the latent representations and observe that efficient reasoning steps naturally cluster into a concentrated region in latent space, while those deviating from this region tend to produce verbose sequences. To leverage this, we keep reasoning focused within this region via a quadratic program which projects deviating hidden states back into the region. Then we propose a novel training-free framework to achieve efficient reasoning that reduces token generation costs without sacrificing performance. Extensive experiments conducted on four models ranging from 1.5B to 14B, and across six benchmarks in math reasoning, coding, and scientific QA, validate the effectiveness of our method, up to a 13.4\% improvement in accuracy while reducing generated tokens by 11.8\% to 52.8\%. Code and models will be publicly available.

Log-Gaussian Cox Processes (LGCP) are essential for modeling spatial point patterns but suffer from high computational costs and intractable likelihoods. We propose VoGCAM, an efficient Variational Voronoi Gaussian Coordinate Ascent Maximization framework for fitting LGCPs. Our approach first approximates the intractable integral in the LGCP likelihood using a flexible Voronoi tessellation, which incorporates observed points as integration nodes. By applying variational Gaussian approximation, we derive an Evidence Lower Bound (ELBO) that admits an explicit and closed-form expression. To optimize this objective, we develop a novel coordinate ascent algorithm that updates parameter blocks via Newton and fixed-point methods. We further enhance scalability by adopting a Nearest Neighbor Gaussian Process (NNGP) prior and utilizing the Woodbury formula to reduce matrix inversion costs. Theoretically, we prove the existence and uniqueness of the optimal solution and establish the convergence of our algorithm. Numerical experiments on synthetic and real-world data demonstrate that VoGCAM offers superior computational efficiency and inferential accuracy over state-of-the-art methods like INLA and VIFRK.

The design of the forward diffusion process is a foundational yet under-theorized aspect of diffusion modeling. Although it determines how data are corrupted and how quickly the forward dynamics approach a reference distribution, forward-process design remains largely governed by canonical defaults rather than a unified analytical framework. We propose Eigenvalue-Guided Control and Analysis (EGCA) for diffusion models, a spectral framework that uses the principal eigenvalue of the infinitesimal generator to analyze and guide forward-process design. Under suitable ergodicity conditions, this eigenvalue governs the exponential convergence rate to stationarity, providing a compact and interpretable summary of forward dynamics. We establish sufficient uniqueness and ergodicity conditions, derive eigenvalue-based convergence bounds, and adapt a practical numerical estimator for the principal eigenvalue. Empirically, we show that the eigenvalue strongly tracks forward convergence speed and serves as an organizing coordinate for the efficiency--quality trade-off across tractable forward-generator families. Experiments across multiple image datasets and diverse diffusion settings, including training formulations, architectures, and samplers, demonstrate that eigenvalue-guided design maintains or improves generation quality while reducing training cost. Beyond a single generator choice, our framework offers a spectral perspective for understanding and comparing forward-process choices in diffusion models.

In this paper, we move beyond the conventional setting of motion in-betweening that relies on structured pose inputs, and instead propose a new paradigm that is conditioned on egocentric (first-person) visual observations, thereby reducing the requirements on both users and upstream systems; we term this task Egocentric Human Motion Bridging. This is motivated by realistic consumer scenarios in VR/AR and robotics, where users typically have access only to raw egocentric imagery (e.g., from wearable devices) without reliable human pose annotations or motion capture signals. Unlike conventional motion interpolation in human pose-conditioned settings or motion generation conditioned on a complete temporal language or image context, the egocentric perspective introduces large viewpoint variations, partial body visibility, and severe temporal ill-posedness due to cross-modal endpoint constraints. To this end, we propose Ego-HMB, a unified diffusion-based model that internalizes both endpoint human pose estimation capability and temporally consistent 3D motion infilling ability within a single model through two training stages. Specifically, we first introduce Egocentric-to-Canonical Pretraining (pretraining stage), where we design a hierarchical curriculum with progressively increasing temporal spans. Then, we present a novel Motion Bridge Diffusion Training with endpoint and physics awareness as a fine-tuning stage, where our bridge diffusion diffuse from neighboring poses to enhance temporal continuity while preserving the global constraints between the two endpoints of a 3D motion sequence. Extensive experiments and comparisons with one-stage baselines (a generative model conditioned directly on two egocentric images) and two-stage baselines (estimation models followed by interpolation models) demonstrate that our approach consistently outperforms these baselines.

Continuous episodic memory is a core capability for autonomous agents operating in dynamic, real-world environments, yet current streaming video benchmarks provide limited tools for diagnosing what models remember and for how long. We introduce Egostream, a diagnostic benchmark for streaming episodic memory evaluation in egocentric vision. \egostream organizes 2,250 curated questions along seven cognitive dimensions: detail, spatial, temporal, event, social, causal, and prospective memory. We introduce the Answer Validity Window (AVW), which specifies the temporal span an answer remains valid as the observed scene evolves. This allows us to expand the questions into 8,528 recall-conditioned evaluations, enabling controlled testing from instant to ultra-long-term recall while separating genuine model forgetting from natural world-state changes. We rigorously establish baseline performance through a unified streaming MLLM framework that compares several state-of-the-art memory-management mechanisms, covering sliding windows, attention sinks, KV-cache pruning, merging, and offloading. Experiments within a unified Qwen3-VL backbone reveal that comparable aggregate accuracies mask starkly different memory profiles. For instance, token pruning preserves fine-grained details and temporal structure significantly better than token merging, while quantized offloading rescues ultra-long-term recall. Ultimately, all mechanisms operate well below real-time (>1s per frame), and top performing methods ceil at about 45\% accuracy, exposing critical gaps in current architectures. Egostream provides the diagnostic testbed needed to close these gaps.

We study secret elicitation: discovering knowledge that an AI possesses but does not explicitly verbalize. As a testbed, we train three families of large language models (LLMs) to possess specific knowledge that they apply downstream but deny knowing when asked directly. For example, in one setting, we train an LLM to generate replies that are consistent with knowing the user is female, while denying this knowledge when asked directly. We then design various black-box and white-box secret elicitation techniques and evaluate them based on whether they can help an LLM auditor successfully guess the secret knowledge. Many of our techniques improve on simple baselines. Our most effective techniques (performing best in all settings) are based on prefill attacks, a black-box technique where the LLM reveals secret knowledge when generating a completion from a predefined prefix. Our white-box techniques based on logit lens and sparse autoencoders (SAEs) also consistently increase the success rate of the LLM auditor, but are less effective. We release our models and code, establishing a public benchmark for evaluating secret elicitation methods.

Fine-tuning LLMs on narrow harmful datasets can induce Emergent Misalignment (EM), where models exhibit misaligned behavior far beyond the fine-tuning distribution. We argue that emergent misalignment can be better understood as a data-mediated transfer phenomenon: harmful fine-tuning examples do not induce uniform behavioral spillover, but interact with the structural properties of the dataset and the difficulty of the tasks relative to the model. Across our experiments, we find that misalignment appears more readily when fine-tuning and evaluation prompts share similar underlying functional structure, when prompts leave more room for coherent harmful completions, and when the target behavior has been more reliably learned by the model. The training pipeline itself also matters: pretraining composition shapes later misalignment. We further study Subliminal Learning (SL), where misalignment is transmitted by fine-tuning on seemingly benign data generated by a harmful teacher. Moving beyond the standard SFT setting, we for the first time compare this transfer under off-policy and on-policy distillation as well, allowing us to separate the roles of the teacher guidance and the training data distribution in transmitting misalignment. Together, these results argue for a data-centric view: Emergent/subliminal misalignment should not be treated as a simple consequence of isolated harmful fine-tuning examples, but as the result of interactions between fine-tuning data structure, pretraining distributions, and training channels.

Text-based counseling is an important interface for AI mental-health support, where transcripts may be used to monitor depression severity and flag sessions requiring timely human review. However, robust PHQ-8 prediction across session regimes remains challenging: fine-tuning-based methods can exploit richer supervision but may generalize poorly under data scarcity, while prompt-based LLM methods are data-efficient but usually treat each transcript holistically and provide limited support for longitudinal context. We study robust depression tracking from counseling transcripts across single-session and multi-session regimes. We introduce LongCounsel-8, a multi-session counseling dataset with session-level PHQ-8 supervision for evaluating repeated-session tracking under partial symptom disclosure and cross-session continuity. We further propose EmoTrack, a PHQ-8 prediction framework that combines LLM-extracted clinical signals with frozen turn-level semantic embeddings and trains symptom-specific predictors over the resulting transcript representation. When prior sessions are available, EmoTrack can further incorporate them through compact cross-session memory. Experiments on LongCounsel-8 and DAIC-WOZ show that EmoTrack achieves a clear gain on the real single-session benchmark, including a 13.5% relative MAE reduction over the strongest DAIC-WOZ baseline, and remains competitive with the strongest longitudinal baseline on LongCounsel-8.

We study methods for simultaneous analysis of many noisy and biased estimates, each paired with an even noisier estimate of its own bias. The analyst's goal is to construct short calibrated intervals for each parameter. The standard debiasing approach, which subtracts the bias estimate from each biased estimate, inflates variance and yields long intervals. In this paper, we propose an empirical Bayes rebiasing strategy that starts from the fully debiased estimates and learns from data how much bias to reintroduce by estimating the unknown bias distribution. We provide convergence rates for the coverage of our intervals when the bias distribution is estimated using nonparametric maximum likelihood. Furthermore, we demonstrate substantial precision gains in prediction-powered inference, including pairwise LLM win-rate evaluations, as well as for inference of direct genetic effects in family-based GWAS.

Passive matching networks are critical components in radio-frequency integrated circuits (RFICs), but their design is bottlenecked by expensive finite-difference-based electromagnetic (EM) simulation. We propose an end-to-end neural framework for free-form RFIC passives that predicts both EM response and downstream design performance under varying frequency and circuit contexts, combining a frequency-conditioned neural simulator with a circuit-conditioned neural solver. To improve design performance prediction, the simulator outputs a neural EM response consisting of both the explicit EM response and a complementary response, together with a dependence regularizer that encourages the two responses to capture orthogonal information. To evaluate the proposed framework, we construct a large-scale dataset of free-form passive layouts, synthesized from both expert-designed contours and random paths. Our 21M-parameter model achieves high accuracy across a broad frequency range, circuit contexts and passive matching settings, with $R^2>0.98$ for almost all targets, while providing orders-of-magnitude faster evaluation than traditional EM simulation.

Latent diffusion models (LDMs) enable high-fidelity synthesis by operating in learned latent spaces. However, training state-of-the-art LDMs requires complex staging: a tokenizer must be trained first, before the diffusion model can be trained in the frozen latent space. We propose UNITE – an autoencoder architecture for unified tokenization and latent diffusion. UNITE consists of a Generative Encoder that serves as both image tokenizer and latent generator via weight sharing. Our key insight is that tokenization and generation can be viewed as the same latent inference problem under different conditioning regimes: tokenization infers latents from fully observed images, whereas generation infers them from noise together with text or class conditioning. Motivated by this, we introduce a single-stage training procedure that jointly optimizes both tasks via two forward passes through the same Generative Encoder. The shared parameters enable gradients to jointly shape the latent space, encouraging a "common latent language". Across image and molecule modalities, UNITE achieves near state-of-the-art performance without adversarial losses or any pretrained encoders (e.g., DINO), reaching FID 2.12 and 1.75 for Base and Large models on ImageNet 256 x 256. Our experiments suggest that unification provides mutual benefits for tokenization and generation. We further analyze the Generative Encoder through the lenses of representation alignment and compression. These results show that single-stage joint training of tokenization and generation from scratch is feasible. Project Page: https://unite-tokenization-generation.netlify.app/

This paper presents EnerGNN, an energy-based framework for exact constrained combinatorial inference. The central idea is to learn an optimization-compatible energy over binary configurations, so that hard constraints and deployment-time restrictions can be enforced by a mixed-integer programming (MIP) layer at inference time. Training shapes the energy landscape using optimality-aware contrastive learning so that high-quality feasible solutions occupy low-energy regions. We evaluate EnerGNN in two settings: direct decision problems, where the learned energy is optimized over the full binary decision vector, and strategic decomposition problems, where it is optimized over complicating binary variables and an exact solver handles the downstream problem. Across Quadratic Knapsack, Combinatorial Auctions, and a personalized medicine supply chain problem, EnerGNN achieves primal gaps below 1\% on most test instances, generalizes to larger problem instances without retraining, and supports zero-shot constraint injection by modifying only the inference MIP.

Structure-informed protein representation learning is essential for effective protein function annotation and $\textit{de novo}$ design. With the rapid progress of protein structure prediction and experimental structure determination, a central challenge in the field is shifting from merely obtaining structural models to reliably leveraging predicted or measured structures for downstream learning. However, both crystal and AlphaFold-like predicted structures can contain structural uncertainty, which can bias local neighborhood construction and make robust protein representation learning challenging. To address these issues, we propose a novel equivariant Transformer-State Space Model (SSM) hybrid framework, termed $E^3$former, designed for efficient and robust protein representation. Our approach uses energy function-based receptive fields to construct proximity graphs that adaptively stabilize neighborhood selection under structural deviations, and incorporates an equivariant high-tensor-elastic selective SSM within the transformer architecture. These components allow the model to adapt to complex geometric interactions and extract structural features with a higher signal-to-noise ratio. Empirical results demonstrate that our model outperforms existing methods in structure-intensive tasks, such as inverse folding and binding site prediction, particularly when using predicted structures, owing to its enhanced tolerance to data deviation and structural uncertainty. Our approach offers a novel perspective for conducting biological function research and drug discovery using imperfect but increasingly available protein structure data. Our code is available on \url{https://anonymous.4open.science/r/E3former-207E}.

We present SetDiff, a geometry-grounded multi-view diffusion framework that enhances novel-view renderings produced by 3D Gaussian Splatting. Our method integrates explicit 3D priors, pixel-aligned coordinate maps and pose-aware Plucker ray embeddings, into a set-based diffusion model capable of jointly processing variable numbers of reference and target views. This formulation enables robust occlusion handling, reduces hallucinations under low-signal conditions, and improves photometric fidelity in visual content restoration. A unified set mixer performs global token-level attention across all input views, supporting scalable multi-camera enhancement while maintaining computational efficiency through latent-space supervision and selective decoding. Extensive experiments on EUVS, Para-Lane, nuScenes, and DL3DV demonstrate significant gains in perceptual fidelity, structural similarity, and robustness under severe extrapolation. SetDiff establishes a state-of-the-art diffusion-based solution for realistic and reliable novel-view synthesis in autonomous driving scenarios.

Accurate time-series forecasting is critical for a wide range of problems involving temporal data. Ensemble modeling is a well-established technique for leveraging multiple predictive models to increase accuracy and robustness, as the performance of a single predictor can be highly variable due to shifts in the underlying data distribution. This paper proposes a new methodology for building robust ensembles of time series forecasting models. Our approach uses Adaptive Robust Optimization (ARO) to construct a linear regression ensemble whose model weights adapt over time. We demonstrate the effectiveness of our method through a series of synthetic experiments and real-world applications, including air pollution management, energy consumption forecasting, and tropical cyclone intensity forecasting. Our results show that our adaptive ensembles outperform the best ensemble member in hindsight by 16-26\% in root mean square error and 14-28\% in conditional value at risk and improve over competitive ensemble techniques.

Policy entropy has emerged as a fundamental measure for understanding and controlling exploration in reinforcement learning with verifiable rewards (RLVR) for LLMs. However, existing entropy-aware methods mainly regulate entropy through global objectives, while the token-level mechanism by which sampled policy updates reshape policy entropy remains underexplored. In this work, we develop a theoretical framework of entropy mechanics in RLVR. Our analysis yields a first-order approximation of entropy change $\Delta \mathcal{H}$, giving rise to entropy polarity, a signed token-level quantity that predicts whether and how much a sampled update expands or contracts entropy. This analysis further reveals a structural asymmetry: Reinforcing frequent high-probability tokens triggers contraction tendencies, whereas expansive tendencies typically require lower-probability samples or stronger distributional correction. Empirically, we show that entropy polarity reliably predicts entropy change, and that positive and negative polarity branches play complementary roles in preserving exploration and strengthening exploitation. Building on these insights, we propose Polarity-Aware Policy Optimization (PAPO), which preserves both polarity branches and implements entropy control through advantage reweighting. With the empirical entropy trajectory as an online phase signal, PAPO adaptively reallocates optimization pressure between entropy-expanding and entropy-contracting updates. Experiments on mathematical reasoning and agentic benchmarks show that PAPO consistently outperforms competitive baselines, while delivering superior training efficiency and substantial reward improvements.

We study semi-offline value estimation, where a small on-policy sample supplements a large off-policy dataset. This regime is common in recommender systems and pre-deployment A/B testing, and the challenge is to use the limited on-policy data to correct systematic bias in a given value predictor. Recent work on Bellman calibration (BC) gives a post-hoc, model-agnostic correction along the predicted value. We propose Error-Guided Bellman Calibration (GBC), which trains an error model on the on-policy sample and calibrates the predictor along this learned error axis. We prove a completeness-free refinement guarantee, a theorem showing that the correction adapts to the intrinsic dimension of the prediction bias, and a finite-sample bound on the value-estimation error that isolates our advantage in the refinement term. Across synthetic, CRM, and D4RL benchmarks, GBC improves over existing calibration baselines.

Foundation models of brain activity aim to learn representations from large-scale neural recordings that generalize across sessions, subjects, species, and downstream tasks. Transformer-based models have emerged as a promising architecture for this goal and begin all with the same first step --- tokenization: transforming sparse binary events into vector representations that downstream networks can consume. In other areas of machine learning, tokenization is its own design problem, distinct from the downstream foundation model. Several mostly implicit tokenization approaches dominate the literature, but their impact on downstream performance has not been systematically evaluated. We propose four explicit desiderata for a neural tokenizer---reconstruction fidelity, downstream utility, cross-instance generalization, and compression---and a framework of metrics spanning seven dimensions, including firing statistics, encoding and decoding, and latent structure and robustness. We assemble a benchmark of ten datasets spanning species, brain areas, recording technologies, and behavioral paradigms, and evaluate a representative set of discrete and continuous tokenizers, both single-neuron and population variants. Across the strategies evaluated, we show how tokenizers trained only on reconstruction can support downstream tasks. Yet reconstruction alone is misleading: reconstruction quality and downstream utility can be sharply dissociated. Our multi-axis framework is a response to this failure mode. These findings highlight tokenization as a critical and underexplored design axis in brain foundation models, and suggest that further progress will require continued systematic study of tokenization strategies.

Physical reasoning is central to agents that step into the physical world. Models now excel at code and math, but their physical reasoning capabilities remain largely untested. Existing physical benchmarks evaluate goal-state matching rather than functional performance. This position paper argues that evaluating physical reasoning in LLM agents requires construction benchmarks, where physics verifies functional outcomes. To operationalize physical reasoning evaluation, we identify five capabilities spanning the full agentic loop from goal to verified artifact. Our audit of eleven physical benchmarks finds only two that combine realistic physics with functional verification, the minimum requirements for testing physical reasoning; both are construction benchmarks. Neither covers all five capabilities, yet frontier models already achieve near-zero success as task complexity increases. Progress depends on the breadth of benchmarks we build. We call on academia, game studios, and industry labs to expand construction benchmarks along both axes, widening capability coverage and deepening physical challenges. Their breadth determines how fully agents move from code and math to reshaping the physical world.

Vision-Language Models (VLMs) have achieved strong performance on static visual understanding benchmarks, yet their ability to act in dynamic environments remains poorly understood. We introduce AtariBench, a controlled benchmark that evaluates VLMs across 20 Atari 2600 games, requiring models to make sequential decisions from high-temporal-resolution gameplay streams. By decoupling fast in-game dynamics from wall-clock inference speed, AtariBench focuses on whether models can perceive rapid visual state changes, reason over spatiotemporal dynamics, and ground actions in the current observation. Our experiments reveal substantial limitations in state-of-the-art VLMs: models often fail to infer object motion and direction, exhibit brittle sequential decision-making, and increasingly over-rely on textual interaction history rather than newly appended visual feedback. We further introduce a separability analysis to identify which in-game scores provide reliable signals of model competence, and use the benchmark to study how memory design and prompting strategies affect performance. Overall, AtariBench exposes a persistent gap between passive visual perception and active decision-making, providing a lightweight and rigorous testbed for developing VLMs that can reason and act in dynamic environments.

Vision-large language models (VLLMs) editing aims to efficiently update model responses while retaining generalization and unrelated knowledge. Recent online editors store previous edits as reusable experts, forming a mixture-of-experts editing memory for scalable VLLM correction. By retrieving relevant experts at inference time, this paradigm avoids repeatedly overwriting model parameters and enables continual reuse of accumulated editing experience. However, existing sample-level residual experts usually encode each edit instance as a whole, entangling target-relevant visual evidence with irrelevant context. This coarse expert construction weakens expert reusability, leads to inaccurate routing, and increases unintended side effects during continual editing. To address this limitation, we reformulate online VLLM editing as evidence guided expert construction and routing. A fine mask identifies edit-relevant visual evidence, from which support-suppress experts are constructed to strengthen target consistent cues and weaken edit conflicting cues. Each expert is stored with routing signals and reliability scores, and later retrieved through joint sparse routing over visual evidence, textual context, and expert quality. Experiments demonstrate consistent improvements in editing reliability, generalization, and interpretability over competitive online editing baselines.

Coding agents are increasingly used as iterative development partners, but most benchmarks still evaluate one specification followed by one final assessment. This fails to capture whether agents can maintain executable artifacts as requirements evolve. We introduce EvoCode-Bench, a benchmark of 26 stateful coding tasks spanning 227 evaluated rounds. Each task preserves the agent's workspace across 5--15 rounds, expresses requirements through observable behavior, and uses cumulative executable tests to check both new requirements and regressions against still-active prior ones. We evaluate 13 coding agents with two complementary metrics: MT@4, a four-attempt fail-stop multi-round score, and SR, a single-round score from a reference-completed prior state. For most agents, SR exceeds MT@4 by 22--40 points, indicating that isolated instruction-following does not imply persistent reliability. The gap also changes model rankings: the highest-SR agent (78.9) ranks only third in persistent execution (44.0 MT@4). Even the strongest agents achieve only \~50\% success on multi-turn metrics, and aggregate pass rate drops below half of round-1 performance by round~5. Failure analysis reveals tier-dependent behavior: weaker agents fail early, while stronger agents survive long enough to expose specification-tracking and regression failures. These results show that single-round and persistent multi-turn coding evaluate different capability dimensions. We release the benchmark data and Harbor multi-turn infrastructure.

The development of native computer-use agents (CUA) represents a significant leap in multimodal AI. However, their potential is currently bottlenecked by the constraints of static data scaling. Existing paradigms relying primarily on passive imitation of static datasets struggle to capture the intricate causal dynamics inherent in long-horizon computer tasks. In this work, we introduce EvoCUA, a native agentic model that integrates data generation and policy optimization into a self-sustaining evolutionary cycle. This approach employs a verifiable synthesis engine to autonomously generate diverse tasks with executable validators, alongside a scalable infrastructure orchestrating tens of thousands of sandbox rollouts for mass experience acquisition. To internalize this experience, our iterative evolving learning strategy reinforces successful routines while transforming failure trajectories into rich supervision through error analysis and self-correction.EvoCUA achieves a 56.7\% success rate on the OSWorld benchmark, establishing a new state-of-the-art for open-weights models across multiple model scales. It significantly outperforms the previous best open-source model, OpenCUA-72B (45.0%), and surpasses leading closed-weights models such as UI-TARS-2 (53.1%), offering a highly parameter-efficient and reproducible alternative to massive closed-weights models. These results demonstrate the generalizability of the evolving paradigm across foundation models of varying scales, establishing a robust and scalable path for advancing native agent capabilities.

High-fidelity diagram creation requires coordinated visual-spatial decisions over semantic topology, visual styling, and geometric layout. Existing methods face a representation gap: pixel-based generation offers limited object-level control, while code-based synthesis provides executable structure at the cost of intuitive manipulation. We introduce EvoDiagram, an agentic framework for editable diagram creation through an intermediate canvas schema that is both machine-actionable and directly manipulable by users. EvoDiagram first constructs a diagram manifest through coordinated structure, style, and layout agents, and then renders the manifest in a diagnostic verification environment. This environment produces objective defect signals and VLM-based critique, enabling localized refinement of the current diagram while also providing evidence for long-term design expertise evolution. The evolution mechanism distills refinement traces into a three-tier design expertise memory, where candidate strategies are linked to the verification evidence that supports them and are promoted into broader guidelines and principles only through repeated cross-context support. We further define CanvasBench, a canvas-recoverable benchmark and evaluation protocol with content, visual, and cognitive dimensions. The current manuscript focuses on the framework, benchmark design, and reporting protocol; empirical results should be added only when the corresponding artifacts, judge traces, and analysis scripts are available for verification. Our code is available at \href{https://anonymous.4open.science/r/evo-diagram/}{https://anonymous.4open.science/r/evo-diagram/}.

Multimodal large language model (MLLM) agents increasingly operate over long histories of dialogues, images, and evolving facts, yet finite context windows and weak persistent memory cause forgetting, temporal inconsistency, and loss of precise past evidence. Existing multimodal memory agents either discard visual evidence by reducing images to captions or treat memory as a static store that is built once and never revised during reasoning; self-evolving agents further accumulate experience via prompting without tight coupling to retrieval and memory revision. Closing this loop with reinforcement learning additionally suffers from credit assignment difficulties over long, multimodal action trajectories. To address these challenges, we propose EvoMM, a reinforced self-evolving multimodal agentic memory framework that unifies memory construction, retrieval, and evolution within a single agentic loop. EvoMM maintains a mutable two-level memory of turn-level short memories and session-level long summaries, organized by a heterogeneous multimodal memory graph with temporal, semantic, image-co-reference, and keyword edges. An iterative PLAN→RETRIEVE→REFLECT→ANSWER policy then revises long memory on the fly and distills structured retrieval experience into a cross-query store for reuse. To optimize this policy, we further introduce PRR-GRPO (Planning-Retrieval-Reflection GRPO), which augments outcome rewards with gated, action-aware process rewards and performs dense credit assignment over a memory-action tree with dual-scale advantage normalization. Experiments on Mem-Gallery, MMLongBench, and LoCoMo show that EvoMM consistently outperforms strong text-only, multimodal, and agentic memory baselines, establishing a training paradigm for memory that reads, revises, and evolves with each query.

Diffusion Transformers (DiTs) have emerged as the state-of-the-art architecture in high-fidelity generative modeling, but their massive computational demands and multi-step inference process limit their deployment in practical scenarios. Post-training quantization (PTQ) offers a promising solution for accelerating inference, but DiTs suffer from severe activation outlier issues, which can lead to catastrophic accuracy loss in integer formats. Although some recent outlier mitigation methods employ orthogonal transformations or diagonal scaling to smooth the activation distribution, they either solve the transformation matrix solely for activations or fail to fully balance the inter-channel variance between activations and weights. Additionally, the construction of the calibration set by taking the weighted average of the activation values for each time step introduces spurious temporal correlations. To address these issues, we propose a joint diagonalization of the activation-weight covariance matrix based on the solution of a generalized eigenvalue problem (GEVP). We also transform the feature fusion of the calibration set's temporal dimension into probabilistic sampling and assembly in the spatial dimension using Uniform Token Splicing (UTS). Specifically, compared to suboptimal baseline methods, our approach reduces the FID by 16.06 and 6.49 on both the W3A4-quantized DiT-XL/2 and PixArt-$\Sigma$ models, respectively. This demonstrates that our method successfully suppresses outliers in activations and weights, maintaining DiT's performance under low-bit quantization conditions. Code is available at the anonymous repository \url{https://anonymous.4open.science/r/JDUS-DiT-74E1}.

Gradient descent is usually studied through external regret, but recent work shows that it controls richer deviations through infinitesimal vector-field guarantees \citep{ahunbay2024first}, semicoarse constraints \citep{ahunbay2025semicoarse}, and weakly convex proximal maps \citep{cai2025proximal}. We identify the finite-deviation invariant behind these guarantees. A feasible deviation \(\phi\colon K\to K\) induces the displacement field \(\Dphi(\xvec)=\xvec-\phi(\xvec)\). Online gradient descent has \(O(\sqrt T)\) regret against every feasible deviation with exact displacement, \(\Dphi=\nabla\Psi\) for a potential function $\Psi$, even when \(\Psi\) is nonconvex. Feasibility handles the projection boundary term, and exactness makes the potential telescope. The condition is sharp for smooth fields on simply connected domains, since any nonzero circulation along a feasible loop can be turned into a bounded cyclic adversary that forces linear regret. Semicoarse affine deviations are the linear-quadratic slice of this class, and proximal deviations are the Moreau-envelope slice. The exact-form class strictly contains weakly convex proximal deviations. For mirror descent, exactness is measured in the mirror geometry through the one-form \(\Dphi(\xvec)^{\transpose}\dd\nabla R(\xvec)\), which subsumes Bregman proximal regret and includes multiplicative deviations for multiplicative-weight updates. In convex games, no-regret with respect to exact-form deviations yields conservative correlated equilibrium, refining coarse and proximal correlated equilibrium.

We study the problem of propagating the mean and covariance of a general multivariate Gaussian distribution through a deep (residual) neural network using layer-by-layer moment matching. We close a longstanding gap by deriving exact moment matching for the probit, GeLU, ReLU (as a limit of GeLU), Heaviside (as a limit of probit), and sine activation functions; for both feedforward and generalized residual layers. On random networks, we find orders-of-magnitude improvements in the KL divergence error metric, up to a millionfold, over popular alternatives. On a variational Bayesian neural network, we show that our method attains hundredfold improvements in KL divergence from Monte Carlo ground truth over a state-of-the-art deterministic inference method. We also give an smooth-distance error bound showing that, under regularity assumptions, moment matching removes the leading low-variance errors and propagates higher-order local accuracy through the layers of a network.

Nonlinear independent component analysis (nICA) is generally unidentifiable due in part to measure-preserving automorphisms (MPAs), which induce indistinguishable latent representations. We show that for non-Gaussian sources with mild regularity assumptions, such MPAs are not Lipschitz, motivating Lipschitz continuity as a structural assumption. Under this condition, we prove global identifiability of nICA within the class of Orthogonal Coordinate Transformations (OCTs) with bounded sources, recovering the true components up to permutation and scaling. Our analysis reduces the learning problem to a linear program over the Birkhoff polytope, and extends to functions with symmetric Jacobians, revealing connections to optimal transport. We introduce OCTNet and SYMNet, normalizing flow architectures that enforce these constraints in practice. Experiments demonstrate strong recovery performance and robustness to model misspecification, outperforming existing approaches.

Multi-modal machine learning has become a dominant paradigm for building large-scale models that integrate information across heterogeneous modalities and achieve unprecedented representational capabilities. However, the multi-modal nature of such models introduces unique challenges for explainability. Specifically, the standard feature attribution task considers only individual feature contributions. While solutions of this kind can be adapted to operate on each modality, such first-order explanations are insufficient to characterize cross-modal feature interactions, which are essential for understanding model behavior involving cross-modal alignment. To bridge the gap, this paper studies higher-order feature interactions in multi-modal settings and proposes Gradient-Estimation-Based Feature Interaction (GEFI). GEFI is derived from the proxy gradient estimation framework, and we establish its theoretical equivalence to the Shapley Interaction Index at arbitrary orders. Beyond the general theory, we instantiate GEFI-PM for CLIP-like models by exploiting their dual-encoder architecture. The resulting formulation enables the efficient estimation of interactions between visual and textual features. Compared to existing methods for revealing cross-modal interactions, which generally operate on fixed image patches, GEFI offers fine-grained pixel-level interactions while maintaining computational efficiency, yielding more expressive results.

Space-time self-similarity (STSS), which captures visual correspondences across frames, provides an effective way to represent temporal dynamics for video understanding.In this work, we explore higher-order STSS and demonstrate how STSSs at different orders reveal distinct aspects of these dynamics. We then introduce the Multi-Order Self-Similarity (MOSS) module, a lightweight neural module designed to learn and integrate multi-order STSS features. It can be applied to a wide range of motion-centric tasks with only marginal computational cost and memory overhead. Extensive experiments on video action recognition, motion-centric video VQA, and robotic tasks consistently demonstrate substantial improvements, validating the broad applicability of MOSS as a general motion modeling module. The source code and checkpoints will be publicly available.

While spatiotemporal deepfake detectors have shown high AUC, our experiments reveal that they remain susceptible to evasion attacks. These models tend to overfit on fragile temporal spectrum cues, rather than learning robust semantic causality. To mitigate this vulnerability, we propose SpInShield, a temporal spectral-invariant defense framework explicitly designed to decouple semantic motion from manipulatable spectral artifacts. We propose a learnable spectral adversary that dynamically synthesizes severe spectral deformations, simulating extreme attack scenarios. By employing a shortcut suppression optimization strategy, SpInShield compels the encoder to extract reliable forensic cues while autonomously purging unstable spectral statistics from the latent space. Experiments show that SpInShield obtains competitive performance on widely used datasets and outperforms the strongest baseline by 21.30\% AUC under simulated amplitude spectral attacks.

Distributional causal inference requires estimating not only average treatment effects but also interventional outcome distributions, including quantiles, tail risks, and policy-dependent uncertainty. As a method for distributional causal inference, generative adversarial network (GAN)-based counterfactual methods are flexible tools for this task. However, these methods have several limitations. First, the objectives of certain techniques do not coincide with the statistical risk of the identifiable causal target, and therefore provide limited theoretical guarantees regarding estimable counterfactual distributions or optimality. Second, they tend to rely on unstable density-based methods, such as density ratio estimation. In this paper, we propose GANICE (GAN for Interventional Conditional Estimation) with several advantages: it (i) clarifies the conditional interventional distribution for each treatment--covariate state as the causal estimation target; (ii) estimates the conditional distribution such that its averaged Wasserstein risk is minimized; (iii) establishes minimax optimality. GANICE achieves these advantages through the introduction of the extended Wasserstein distance, the incorporation of a cellwise critic in its dual, and an optimality proof based on Besov space theory. Our experiments demonstrate that GANICE consistently outperforms existing methods.

Air quality prediction at unmonitored locations is challenging because historical AQI observations are unavailable, while conventional spatiotemporal models largely rely on target histories to learn spatial dependencies. Meanwhile, widely available weather covariates provide useful station-specific evidence for target-absent inference. In this paper, we propose \textbf{ExtrapAir}, a lightweight weather-bridged framework for air quality inference at unmonitored locations. ExtrapAir preserves variable-specific weather patterns through per-variable weather encoding and learns adaptive weather and AQI correlations for spatial transfer. To support robust inference, we introduce \textbf{Activation Attention with Prior Inductive Biases}, which enables non-competitive spatial aggregation, incorporates geographic and semantic priors for station-pair guidance, and avoids full station-to-station attention computation for scalability. Extensive experiments on five real-world air-quality datasets against \textbf{32} baselines show that ExtrapAir reduces MAE by up to \textbf{15.24\%} overall and \textbf{16.01\%} on unmonitored stations, while cutting training time by \textbf{84.9\%} and memory usage by \textbf{76.23\%}, achieving a strong accuracy--efficiency trade-off.

Linear interpolation between fine-tuned checkpoints has been shown to trace the Pareto front between competing objectives, but whether extrapolative weight averaging can extend such frontiers to new checkpoints useful at inference time, without additional RL training, remains unclear. We study this question in RL for competitive programming, where hidden unit tests under time and memory limits enforce both functional correctness and computational efficiency. Starting from a shared initialization, we train checkpoints under nested unit-test coverage: low-coverage rewards require passing smaller-input tests, while high-coverage rewards require passing progressively larger tests up to the full suite. This sweep reveals the emergence of a correctness–efficiency frontier: on hard problems, higher-coverage reward reduces optimization failures but increases correctness failures, leaving solve rate nearly unchanged. Interpolation between low- and high-coverage checkpoints recovers this frontier, while extrapolation extends it beyond the trained endpoints. Both the frontier and its extrapolative continuation appear across three inference settings, pure reasoning, tool use, and agentic coding, and across two model scales, 32B and 7B. At the problem level, moving along the frontier changes which problems are solved, making extrapolated checkpoints complementary policies in inference-time scaling: ensembles with extrapolative weight averaging broaden coverage and improve pass@250 on LCB/hard by 3.3\% over the best single checkpoint at matched sample budget. These results show that nested unit-test coverage in code RL induces a frontier that extrapolative weight averaging can navigate, extend, and exploit.

Existing image watermarking methods typically entangle visual content with spatial structure, making them highly sensitive to geometric transformations and alignment discrepancies, especially in the presence of deepfake manipulations. In this paper, we propose a semantic-driven facial watermarking framework for robust identity recovery. The key idea is to decouple identity-related semantic information from spatial layout and encode it into a compact and spatially robust semantic representation. Specifically, we decompose a face into semantic components and aggregate deep features into component-wise latent representations, which are quantized via independent codebooks and converted into a compact bitstream for embedding. After decoding, the embedded semantic information is recovered and used to reconstruct identity-consistent facial content, even under slight geometric distortions and tampering. Experiments on CelebA-HQ and FFHQ demonstrate that our method significantly outperforms existing watermarking approaches in terms of reconstruction quality, identity preservation, and retrieval accuracy under both photometric and geometric attacks, validating the effectiveness of semantic component-wise encoding for reliable face recovery.

Diffusion Transformers (DiTs) have established themselves as the preeminent paradigm for high-fidelity generative modeling across various modalities, yet their practical deployment is hindered by the severe computational overhead of iterative sampling. While recent feature-caching methods mitigate this latency through feature reuse or forecasting based on integer-order operators, they remain limited in adapting to the non-stationary and highly curved dynamics of DiT latent trajectories. In this work, we reveal that DiT feature evolution exhibits anomalous diffusion characteristics and a certain degree of spatiotemporal heterogeneity, rendering static, integer-order forecasting operators inadequate and prone to error accumulation. To address this, we propose Fractional Anomalous Dynamics Extrapolation (FADE), a unified framework for training-free diffusion transformers acceleration. FADE models DiT feature evolution via anomalous diffusion and employs a Mittag--Leffler fractional operator to better adapt to high-curvature and non-linear latent trajectories of DiTs. Leveraging the empirical observation that these complex dynamics maintain cross-sample stability, we further construct a Look-Up Table with multiple dimensions through offline profiling. This enables zero-overhead, spatiotemporally adaptive parameter retrieval during online inference. Extensive evaluations across diverse architectures and modalities demonstrate that FADE achieves state-of-the-art acceleration, preserving exceptional generation fidelity under aggressive skip intervals with negligible latency overhead. Our code is anonymously available at https://anonymous.4open.science/status/FADE-8CB5.

Continuous-time models are a natural choice for irregular and asynchronous data. A central design choice is how to embed discrete observations into continuous time. Interpolation- and imputation-based embeddings reconstruct a continuous observation path, making the model sensitive to the choice of reconstruction. We show that this reconstruction step is unnecessary. On compact sets, universality for continuous functionals of paths transfers to universality for continuous functionals of discrete observation streams under any continuous and injective embedding. Guided by this result, and building on the rectilinear control path for Neural Controlled Differential Equations (NCDEs), we introduce a continuous and injective embedding for Log-NCDEs, a universal class of continuous-time models. This embedding records observations as increments and composes them over arbitrary query intervals to form log-signatures, giving interval-level summaries of a faithful embedding of the observed data. This avoids interpolation of observed variables, supports online computation, and allows prediction on output grids independent of the input sampling times. Experiments on synthetic controlled dynamics and real-world time-series datasets show that the representation is accurate, efficient, and robust to irregular, asynchronous, and sparse observations.

Chains of thought (CoTs) have become central in interpreting and auditing behaviors of large language models. Yet growing evidence suggests that these traces often fail to faithfully represent the computations behind a model's predictions. Several faithfulness metrics have been proposed, but whether they indeed measure faithfulness remains unknown. Answering this requires ground-truth labels, which are hard to obtain since internal computations are not directly observable. Consequently, most works proposing metrics report only absolute scores or comparisons to prior metrics, and the few existing benchmarks rely on proxies like plausibility or importance, properties orthogonal to faithfulness that can mislead about whether a CoT can be trusted. We address this challenge by constructing tasks whose outputs reveal which intermediate computations must have produced them, and developing an automated labeling pipeline that yields ground-truth faithfulness labels at both the step and CoT level. Building on this methodology, we present BonaFide, a benchmark of 3,066 labeled CoTs across 13 tasks and 10 models, and use it to conduct the first systematic evaluation of prominent faithfulness metrics. Our experiments show that most metrics perform near chance, exhibit strong prediction biases and degrade on longer CoTs. The best metric reaches only 0.71 AUROC at the CoT level while another reaches 0.59 at the step level, with neither transferring across settings, while entailing prohibitively high computational cost. Our results expose fundamental gaps in current faithfulness evaluation and call for the development of more reliable and efficient metrics.

Large language models can now generate complex, multi-step mathematical proofs, but reliably determining their correctness and localizing early logical errors remains a critical challenge. Existing evaluation approaches largely depend on model-based natural-language judgments, which often overlook local reasoning gaps. While formal theorem provers like Lean offer a path to rigorous verification, using them to evaluate informal text requires solving locality and semantic mismatches: a prover might bypass a local flaw by proving an overly broad target, or validate an auto-formalized statement that drifts from the original mathematical intent. To address this, we introduce FaithSieve, a Lean-assisted framework for fine-grained evaluation of natural-language mathematical proofs. FaithSieve decomposes coarse proof steps into local reasoning units, extracts typed proof obligations, and verifies them through a formal evaluation agent. Formal validation is gated by semantic alignment scoring, so Lean evidence is incorporated only when the formal statement faithfully preserves the context, objects, and logical form of the original claim. We construct two expert-verified datasets, ProofLoc-Olympiad and ProofLoc-University, to benchmark first-error localization. On the 350-problem Olympiad dataset, FaithSieve using a GPT-5.4 backbone achieves 81.43\% exact first-error accuracy, outperforming the direct-judging baseline of 72.29\%. Furthermore, on the 200-problem ProofLoc-University benchmark spanning six advanced domains, FaithSieve reaches 84.5\% exact accuracy, compared to 75.0\% for the direct judge. Our work demonstrates that decomposing proofs into fine-grained units and grounding them with faithful formal evidence significantly improves reliable evaluation of natural-language reasoning.

We propose FA-LAM, a focus-aware large avatar model for one-shot animatable Gaussian head creation, while simultaneously enabling static 3D and dynamic 4D full-head recovery. The core of our method lies in a thorough analysis of the attention mechanisms and the entangled reconstruction and animation training pipeline adopted by prior state-of-the-art approaches. Our analysis identifies two main factors that compromise the quality of 3D full-head generation: (1) incorrect and noisy attention activations, and (2) conflicts between the tasks of reconstruction and animation. To address the first issue, we introduce a symmetric and semantic attention regularization strategy that leverages the inherent semantics and structural symmetry of human heads. To disentangle the objectives of reconstruction and animation, we develop a novel dual-phase training pipeline that separates the model's capabilities for large-view hallucination and animation into distinct modules. Moreover, we enhance our model to support multi-view and streaming 4D reconstruction in an efficient and memory-friendly manner through a core autoregressive modification with tailored visibility-aware token fusion. Collectively, these innovations enable FA-LAM to reconstruct animatable Gaussian full heads with superior quality, particularly in fine facial regions and large viewing angles.

Gradient descent has been of particular interest in modern machine learning beyond sole focus on optimization because some specific structures emerge through the optimization dynamics. Such behaviors result from optimization even though the learning objective does not explicitly encode the target structure, collectively called implicit bias, often preventing overparametrized models from fitting to spurious patterns. A typical instance is the max-margin implicit bias of a linear classifier, widely established for exponentially tailed loss functions. Even after having a given dataset separated, the parameter vector continues to evolve towards the max-margin direction asymptotically along the gradient descent dynamics. This phenomenon corroborates a frequent empirical observation of ``train longer, generalize better.'' However, the max-margin convergence is an asymptotic phenomenon, and what is worse, this asymptotic convergence rate is significantly slower than convergence in optimization. Even so, the parameter vector along gradient descent dynamics commonly correlates with the max-margin direction positively (though not exactly) within considerably fewer iterations than the asymptotic rate. By shedding another light on this classical yet profound problem, this work aims to understand the mechanism of this early-stage alignment phenomenon. Our theoretical results demonstrate that the parameter vector weakly aligns with the max-margin direction within $O(\exp(\exp(-\delta)))$ iterations, where $\delta>0$ is the permissible alignment error, which is shown to be tight. By tracking the radial and tangential flows, our proof operates on the alignment dynamics directly with dataset geometry and gets rid of the asymptotic expansion, which is a key insight to enabling faster weak alignment.

Clifford algebra is becoming increasingly common in machine learning, e.g., in neural operators, equivariant transformers, and structured linear layers. The main appeal is that scalars, vectors, oriented planes, volumes, and higher-grade quantities live together in a single algebraic space. In this setting, rotations are represented by \emph{rotors}, which are parametrized by bivectors, with only $\binom{n}{2}$ degrees of freedom. Most Clifford algebra-based models work only with small algebras, in part because of a bottleneck in the sandwich product, $x \mapsto r x r^\dagger$, a basic operation for applying rotations to multivectors. It is evaluated through geometric products or via a full action matrix, both of which scale poorly with the $2^n$-dimensional Clifford basis. We ask whether the $\binom{n}{2}$-parameter description can be carried all the way through application, and show that the answer is yes. The sandwich action decomposes grade-wise into exterior powers of the induced vector rotation. We factor this vector rotation into Givens rotations and apply their lifted actions as sparse updates on pairs of blades. The resulting algorithm never needs to form the rotor coefficients or the $2^n \times 2^n$ action matrix. Empirically, our algorithm achieves up to $64 \times$ wall-clock and $35\times$ memory improvements over existing rotor-sandwich implementations, enabling rotor-based layers at scales previously out of reach.

Face recognition has achieved strong performance, but further gains often require much larger labeled datasets, making annotation costly. This motivates the use of unlabeled data, where face clustering is important for applications such as annotation and recognition. Recent GCN-based methods have shown promising results on affinity graphs, but balancing clustering quality and scalability on large-scale data remains challenging. To solve this problem, we propose FC-DGCN, a face clustering framework based on deep residual GCNs. The proposed method formulates clustering on an affinity graph through two complementary tasks: node confidence estimation and edge connectivity estimation. Specifically, we develop DGCN-V to estimate node confidence on the global graph and DGCN-E to estimate edge connectivity on local subgraphs. To better characterize the local cluster structure, we further revise the confidence-estimation target. Based on these two modules, FC-DGCN links each node to a higher-confidence neighbor with strong predicted connectivity, which naturally induces identity-consistent clusters. Experiments on MS-Celeb-1M show that FC-DGCN outperforms competitive baselines in Pairwise F-score and BCubed F-score while remaining computationally practical. Moreover, the clusters generated by FC-DGCN provide effective pseudo labels for low-label face recognition, leading to substantial improvements on MegaFace and IJB-A.

Policy gradient methods serve as a cornerstone of reinforcement learning (RL), yet their extension to safe RL, where policies must strictly satisfy safety constraints, remains challenging. While existing methods enforce constraints in every policy update, we demonstrate that this is unnecessarily conservative. Instead, each update only needs to progressively expand the feasible region while improving the value function. Our proposed algorithm, namely feasible policy optimization (FPO), simultaneously achieves both objectives by solving a region-wise policy optimization problem. Specifically, FPO maximizes the value function inside the feasible region and minimizes the feasibility function outside it. We prove that these two sub-problems share a common optimal solution, which is obtained based on a tight bound we derive on the constraint decay function. Extensive experiments on the Safety-Gymnasium benchmark show that FPO achieves excellent constraint satisfaction while maintaining competitive task performance, striking a favorable balance between safety and return compared to state-of-the-art safe RL algorithms.

Model merging combines task experts into one model and avoids joint training, retraining, or deploying many expert models, but the merged model often still underperforms task experts. We study this performance gap through feature drift, the difference between features produced by the merged model and by the expert on the same input. Our theory decomposes this drift into upstream propagation and local mismatch, tracks how it propagates and combines through later layers in forward order, and links final feature drift to output drift. This view motivates FeatCal, which uses a small calibration set to calibrate the merged model weights layer by layer in forward order, reducing feature drift while staying close to merged weights and preserving the benefits of model merging. FeatCal uses an efficient closed-form solution to update model weights, with no gradient descent, iterative optimization, or extra modules. On the main CLIP and GLUE benchmarks, FeatCal beats Surgery and ProbSurgery, the closest post-merging calibration baselines: 85.5% vs. 77.0%/78.8% on CLIP-ViT-B/32 Task Arithmetic (TA) and 85.2% vs. 83.7%/82.2% on FLAN-T5-base GLUE. On CLIP-ViT-B/32, 8 examples per task reach 82.9%, and 256 examples per task take 53 seconds, about 4x faster than both baselines, showing better sample efficiency and lower calibration cost.

Fully Homomorphic Encryption (FHE) enables non-interactive, privacy-preserving transformer fine-tuning, protecting highly sensitive data as inputs to large language models. Prior work has predominantly pursued efficiency gains through architectural refinements, yet overall performance remains bounded by the substantial cost inherent in encrypted computation. To address these issues, we target the bottleneck at the ciphertext operation level. Concretely, we propose a novel packing strategy that supports batching over a flexible number of inputs, making it well-suited for fine-tuning workloads. In addition, we design operation-specific matrix evaluation algorithms that significantly reduce (i) the number of ciphertext rotations and relinearizations, and (ii) the number of bootstrapping invocations. Additionally, we exploit functional bootstrapping to fold the evaluation of element-wise activation functions directly into the bootstrapping step, thereby reducing overall computational complexity. As a result, our framework achieves a $30\times$ reduction in ciphertext rotations and relinearizations across all matrix operations, as well as a $2.5\times$ reduction in FHE bootstrapping invocations over the entire computation flow, compared to the latest state-of-the-art work. It is worth noting that our system completes fine-tuning of a 2-layer BERT-style model with a batch size of 16 in 650 seconds, achieving a $8\times$ speedup over the state-of-the-art work.

Low-Rank Adaptation (LoRA) is a popular parameter-efficient fine-tuning method for Large Language Models (LLMs), and practical systems usually need to serve many concurrent LoRA adapters to process different tasks. This is challenging due to the inherent tension between \textit{fairness} and \textit{efficiency}, i.e., achieving fairness requires frequent switching among the adapters such that they have similar service quality, while good efficiency requires to minimize adapter switching to reduce memory movement overheads. Existing schedulers can only achieve a fixed trade-off between fairness and efficiency, and their trade-offs are often suboptimal. To tackle this problem, we propose \textbf{F}airness-aware \textbf{E}fficient \textbf{L}oRA \textbf{P}riority \textbf{S}cheduling (FELPS), which explicitly decomposes fairness and efficiency considerations. In particular, for fairness, FELPS adapts the classic proportional fairness and ensures that the amount of service received by each adapter is proportional to its workload. For efficiency, FELPS prioritizes adapters that are already loaded to GPU memory. To conduct scheduling, FELPS combines to the fairness and efficiency terms with an adjustable weight to achieve arbitrary trade-offs between fairness and efficiency. Evaluations on real workloads show that FELPS achieves a superior trade-off between fairness and efficiency than all baselines.

Vision-language models (VLMs) have achieved remarkable progress. However, processing high-resolution images and long videos generates a massive number of vision tokens, creating a severe inference bottleneck due to the quadratic complexity of attention mechanisms. Previous efforts to mitigate this issue have primarily focused on accelerating the prefill stage by reducing the number of vision tokens, often yielding limited practical speedups while neglecting the increasingly time-consuming decoding stage when the generated response is long. In addition, existing approaches heavily rely on attention maps or similarity maps that are not compatible with efficient attention implementations. To comprehensively address these limitations, we first propose **FlashPruner**, a lightweight module that selects the most helpful vision tokens without relying on attention maps, significantly reducing prefill time. For the decoding stage, we adopt self-speculative decoding and theoretically prove that on-policy distillation guarantees a lower bound for the speculative acceptance rate. Guided by this formulation, we introduce **LoFT**, an on-policy distillation method that effectively trains a highly reliable and efficient draft model including FlashPruner and early layers of the language model. Extensive experiments demonstrate that after training with LoFT, FlashPruner achieves substantial inference speedups while preserving task performance. Notably, on MVBench, FlashPruner improves the time-to-first-token by 1.5$\times$ and decoding speed by 2$\times$, maintaining 96.8\% relative accuracy with only 10\% vision tokens.

Monocular geometry estimation has recently achieved impressive performance across diverse scenes. However, state-of-the-art models still face notable distortion in local 3D structure, especially in fine details, like thin structure and small objects. We attribute this limitation to an architectural mismatch: most current models decode 3D geometry within a 2D parameterization, where feature interactions are governed by image-plane proximity rather than true 3D spatial relationships. This inadvertently mixes features from geometrically distant surfaces, resulting in over-smoothed geometry particularly around thin or elongated structure. In this paper, we propose a fine-detail monocular geometry estimation with Self-Guided Sparse 3D Refinement (SSR) that lifts monocular geometry modeling from 2D image space to 3D space for high-fidelity metric-scale point map. Our model lifts the coarse point map from a foundation base model onto a sparse voxel shell and refines it via SSR. The SSR employs sparse convolutions that aggregate features based on 3D spatial locality, avoiding feature mixing across depth discontinuities. Extensive experiments on diverse datasets demonstrate that our method significantly outperforms existing approaches in recovering fine detailed 3D geometry across both quantitative metrics and qualitative visualizations.

Fine-tuning is widely used to align large language models (LLMs) with desired behavioral norms such as safety, yet it is typically evaluated using aggregate metrics that assume uniform improvements across the model. However, fine-tuning operates within a constrained parameter space, and its entangled effects on localized knowledge remain poorly understood. To address this, we investigate fine-tuning at the level of individual knowledge concepts using a controlled framework with representative concepts and paired safe/unsafe samples under varying data compositions. By jointly analyzing behavioral safety via model outputs and internal knowledge accessibility via perplexity, we obtain a unified view of how knowledge is modified and expressed. Our results reveal that fine-tuning does not uniformly improve safety. Instead, it (1) exhibits strong concept-level heterogeneity, including persistent low-safety concepts and robustly safe concepts with clear domain specificity; (2) improves safe–unsafe discrimination on average, but coexists with familiarity on both safe and unsafe content or even loses previously learned safe knowledge; and (3) decouples internal knowledge from external behavior, as changes in familiarity do not reliably translate into safer outputs, especially due to early-stage disruption of existing safety structures. Our findings position consistency as a distinct objective beyond conventional aggregate safety and demonstrate that reducing unsafe knowledge exposure under uniform concept-level coverage actively homogenizes safety distributions across model knowledge, offering a principled path toward alignment that is not merely effective on average, but uniformly reliable, predictable, and robust across knowledge concepts.

Neuroscientists have recently turned to intracranial brain recording methods, like electrocorticography (ECoG), for human experiments because of the fine spatial and temporal resolution that they afford. Models trained on this data, however, are fundamentally restricted by the patient populations that can receive the implants necessary for recording. We propose using non-invasive fMRI to bridge the gap in training data. Using spoken language representations fine-tuned on fMRI, we build encoding models of ECoG. These representations showed improved prediction performance in ECoG, even though the temporal resolution of fMRI is two orders of magnitude worse. Prediction improved in frequency bands well beyond what is directly measured in fMRI. Next, to test the procedure's generalization ability, we fine-tuned models on fMRI responses that were temporally downsampled by a factor of 2. Despite the loss in resolution, these models were able to predict fMRI and ECoG responses at levels comparable to the original fMRI-tuned models. Finally, we showed that ECoG performance steadily scales with the amount of fMRI-tuning data. Our results show that "slow" data like fMRI can be a valuable resource for building better models of "fast" brain data like ECoG. In the future, integrating across multiple recording methods may further improve performance in other applications, like decoding.

Large language models (LLMs) are increasingly deployed in financial research workflows, where their role is evolving from single-model assistance for human analysts toward autonomous collaboration among multiple agents. Yet recent real-world deployments reveal persistent failures such as factual errors, numerical inconsistencies, and shallow analysis, which can distort assessments of corporate fundamentals and trigger severe economic losses. While existing benchmarks have begun to evaluate such failures, they score all aspects of the generated analysis in one pass, failing to distinguish whether a model fails at foundational stages like auditing and correction, or underperforms at generating research-grade insights. Consequently, it obscures capability bottlenecks and the specialized strengths essential for multi-agent role assignment. To address these gaps, we introduce FinReasoning, a hierarchical benchmark that decomposes the core capabilities of financial research into semantic consistency, data alignment, and deep insight. We further propose a fine-grained evaluation framework that strengthens hallucination-correction assessment and incorporates a 12-indicator rubric for core analytical skills. FinReasoning reveals clear capability stratification across model types. Closed-source models (like Doubao-Seed-1.8) perform strongly overall and are better suited for core reasoning agents in multi-agent financial systems; open-source general models (like Qwen3-235B) show clear capability divergence and consistently underperform on Semantic Consistency, making them less suited for quality-sensitive generation tasks; financial-domain models (like Fin-R1) generate moderate insights but lack foundational auditing skills, requiring upstream quality control for domain-specific analysis. Our work provides a reliable evaluation foundation for multi-agent financial application systems and has already been deployed in pilot tests across several real-world scenarios. The resource is available at https://anonymous.4open.science/r/FinReasoning.

With the rapid progress of Multimodal Large Language Models (MLLMs), unified MLLMs that jointly perform image understanding and generation have advanced significantly. However, despite the inherent reasoning capabilities of unified MLLMs for self-reflection and self-refinement, their use in text-to-image generation remains largely underexplored. Meanwhile, existing multimodal reasoning–based image generation methods mostly rely on prompt augmentation or holistic image–text alignment judgments, without fine-grained reflection and refinement of detailed prompt attributes, leading to limited fine-grained control. To address this limitation, we propose FiRe, a Fine-grained Multimodal Reasoning method for enhanced image generation by MLLM. In specific, FiRe performs a fine-grained multi-step reasoning by first decomposing the prompt into key visual requirements and then self-judging their satisfaction in the generated image, followed by localized refinement according to self-generated precise feedback. In addition, to further strengthen the MLLM's multimodal reasoning ability, we introduce FiRe-GRPO, a reinforcement learning method tailored to FiRe. Since standard Group Relative Policy Optimization (GRPO) suffers from sparse, outcome-based rewards in multi-step reasoning, we formulate our reasoning process as a step-level decision-making problem, design step-specific rewards, and compute step-level advantages for granular credit assignment within GRPO. Extensive experiments demonstrate that FiRe consistently outperforms competitive text-to-image baselines, including existing reasoning-based methods, with particularly substantial gains on compositional text-to-image benchmarks.

We introduce First-Order Trajectory Matching (FTM), a surrogate-modeling method that learns the first-order local transport of probability mass from trajectories of stochastic systems. By matching the symmetric first-order motion of trajectories, FTM learns the probability current velocity, whose flow preserves time marginals to match ensemble averages, while also capturing current-like trajectory quantities such as fluxes, circulations, and barrier-crossing currents. FTM learns the current velocity directly from trajectories, avoiding drift, diffusion, and score estimation. Our stability analysis separates discretization error from sampling variance and shows that the one-step simulation-free FTM loss is stable when temporal resolution and sample size are properly balanced. Across stochastic dynamical systems and PDE examples, we empirically demonstrate that FTM provides trajectory-aware ensemble predictions at low, deterministic-rollout cost.

Multivariate time-series forecasting requires reliable cross-series interaction, yet many attention- and mixing-based models couple heterogeneous variables without explicit entry-stage alignment. Such uncalibrated early interaction can lead to poorly conditioned optimization and degraded generalization when cross-series statistics drift. We propose First-layer Statistical Calibration (FISC), a simple framework that imposes statistical calibration constraints only on the first cross-series coupling layer while leaving deeper coupling layers learnable without repeated calibration. FISC estimates training-set cross-series statistics, converts them into a statistical affinity matrix, and convexly fuses this matrix with learnable coupling parameters. A row-wise simplex projection then yields a nonnegative row-stochastic mixing operator, providing bounded and interpretable entry-layer interaction. By restricting calibration to the first layer, FISC calibrates the entrance interaction without repeatedly constraining deeper feature transformations. To reduce temporal redundancy, FISC further combines a time-domain branch with a data-adaptive orthogonal-domain branch derived from training-time temporal statistics. Experiments on standard long- and short-term forecasting benchmarks show that FISC improves forecasting accuracy while maintaining favorable efficiency. Ablations and non-stationarity analyses further show that first-layer-only calibration consistently outperforms no calibration, last-layer calibration, and all-layer calibration, supporting a simple principle for robust cross-series learning: constrain the entrance interaction with training-set statistics, while letting deeper representations remain data-adaptive.

Mixed-Integer Linear Programming (MILP) is a fundamental tool for combinatorial optimization with extensive real-world applications. A central challenge lies in designing efficient MILP formulations. Large Language Models (LLMs) offer new opportunities to automate the modeling process, from deriving formulations to strengthening them. To ensure correctness, we need robust methods to compare formulations. However, existing approaches evaluate formulations numerically and fail to reason about the behavior on general problem instances. We resolve this limitation by introducing a constructive notion of MILP reformulation that can be formalized in Lean and machine-checked. We develop FLARE, a method that uses an LLM-based agent and the Lean proof assistant to verify proposed reformulations against a reference. To evaluate our approach, we introduce FormulationBench, a challenging dataset of 20 problems and 116 formulations. FLARE outperforms existing methods with 96.9% accuracy. For cases where formal guarantees aren't necessary, we introduce FLARE-NL , an LLM proxy that achieves 99.3% accuracy. These methods enable reliable verification in automated optimization modeling.

3D geometric understanding is undergoing a major transition where rigid, deterministic, optimization-based pipelines are being replaced by large-scale, deep learning models. While recent methods achieve impressive performance on tasks such as multi-view 3D reconstruction and camera pose estimation, they suffer from two critical limitations: $\textbf{(i):}$ current state-of-the-art (SOTA) models require industry-level compute and training corpora, severely limiting accessibility, and $\textbf{(ii):}$ recent approaches remain constrained to modifying existing pretrained transformer models, since substantially changing their architecture would require retraining from scratch. We instead study a complementary direction: training an attention-free state space model for dense 3D understanding. We introduce $\textbf{FlashSSM-3D}$, the first 3D reconstruction state space model capable of performing accurate geometric reasoning over hundreds of images with linear computational complexity. However, training such a model from scratch with competitive performance is impractical under typical academic compute constraints. To address this, we train FlashSSM-3D by distilling knowledge from VGGT, avoiding the need to learn 3D geometric priors entirely from scratch. However, standard distillation losses are poorly suited for dense 3D reconstruction, where preserving real-world spatial structure is essential. We therefore propose a $\textbf{geometry-aware knowledge distillation}$ method specifically tailored for large-scale 3D reconstruction models. Our method introduces two novel losses: a Procrustes-based alignment loss that enforces global structural consistency between student and teacher 3D point predictions, and a camera reprojection loss that directly supervises view-consistent camera geometry between teacher and student. Extensive experiments across multi-view 3D reconstruction and camera pose estimation benchmarks demonstrate that FlashSSM-3D achieves competitive performance while requiring only $\sim25\\%$ of the training data and less than $10\\%$ of the training compute used by VGGT. At inference time, FlashSSM-3D benefits from the practical efficiency of state space models, reconstructing scenes from 500+ views in under 10 seconds while maintaining competitive accuracy.

The Lipschitz constant of neural networks is widely used in robustness, control, generalization bounds, and training stability. As modern neural networks continue to scale, existing methods for Lipschitz estimation are either very time-consuming or overly loose. In this work, we focus on both fast and accurate Lipschitz estimation for large feedforward neural networks. We first reformulate the well-known semidefinite program (SDP) framework into an explicit recursive form, and introduce a key theoretical result, bound backpropagation: the SDP objective has upper and lower bounds that can be propagated backward along the recursive computation graph and depend only on partial variables. Leveraging this property, we decompose the original SDP into a sequence of small subproblems, each minimizing a combination of upper and lower bounds over partial variables. We prove that each subproblem is one-dimensional and convex, enabling fast solution via the secant method. Finally, we present the full FLIP algorithm and its complexity, yielding fast and tight Lipschitz estimation for large feedforward networks. Experiments on both randomly generated and trained networks show that FLIP improves speed and tightness by several orders of magnitude over prior methods, giving the tightest Lipschitz estimates for 100M-parameter networks in 3 seconds.

Recent work has demonstrated that online reinforcement learning (RL) can substantially improve the quality and alignment of flow matching models for image and video generation. Methods such as Flow-GRPO and CPS cast the denoising process as a Markov Decision Process and apply PPO-style ratio clipping to enforce a trust region. However, we argue that ratio clipping is structurally ill-suited for flow models: the probability ratio between new and old policies is a noisy, single-sample estimate of the true policy divergence, leading to over-constraining in some regions of the trajectory and under-constraining in others. We propose Flow-DPPO (Flow Divergence Proximal Policy Optimization), which replaces ratio clipping with a divergence proximal constraint. A key observation is that the per-step policy in flow models is Gaussian, enabling exact and cheap computation of the KL divergence between old and new policies. Flow-DPPO employs an asymmetric divergence mask that blocks gradient updates only when they simultaneously move away from the trusted region and violate the divergence threshold. Experiments show that Flow-DPPO achieves higher rewards with better KL-proximal efficiency, alleviates catastrophic forgetting, promotes balanced multi-objective optimization, and enables stable multi-epoch training where ratio clipping degrades.

Recent vision-language-action (VLA) models can generate plausible end-effector motions, yet they often fail in long-horizon, contact-rich tasks because the underlying hand-object interaction (HOI) structure is not explicitly represented. An embodiment-agnostic interaction representation that captures this structure would make manipulation behaviors easier to validate and transfer across robots. We propose FlowHOI, a two-stage flow-matching framework that generates semantically grounded, temporally coherent HOI sequences, comprising hand poses, object poses, and hand-object contact states, conditioned on an egocentric observation, a language instruction, and a 3D Gaussian splatting (3DGS) scene reconstruction. We decouple geometry-centric grasping from semantics-centric manipulation, conditioning the latter on compact 3D scene tokens and employing a motion-text alignment loss to semantically ground the generated interactions in both the physical scene layout and the language instruction. To address the scarcity of high-fidelity HOI supervision, we introduce a reconstruction pipeline that recovers aligned hand-object trajectories and meshes from large-scale egocentric videos, yielding an HOI prior for robust generation. Across the GRAB and HOT3D benchmarks, FlowHOI achieves the highest action recognition accuracy and a 1.7$\times$ higher physics-simulation success rate on GRAB than the strongest diffusion-based baseline, while delivering a 40$\times$ inference speedup. We further demonstrate real-robot execution across four contact-rich, dexterous manipulation tasks, highlighting the effectiveness of retargeting our generated HOI representations to robotic dexterous hands, achieving significantly higher success rates than prior methods due to improved physical plausibility and semantic alignment.

Language models based on discrete diffusion have attracted widespread interest for their potential to provide faster generation than autoregressive models. Despite their promise, these models typically produce samples whose quality sharply degrades in the few-step regime, preventing a dramatic speedup in practice. Here, we show that language models based on continuous flows over one-hot token embeddings can outperform discrete diffusion in both quality and speed. Importantly, our continuous formulation defines a unique flow map that can be learned directly for efficient few-step inference, a structure we show is unavailable to discrete methods. In this setting, we show that both the flow and its associated flow map can be learned with simple cross-entropy objectives that respect the simplex geometry of the data, and we identify three distinct choices for flow map distillation whose performance we compare in practice. Using these insights, we build a flow language model (FLM), a continuous flow that matches state-of-the-art discrete diffusion baselines on the One Billion Words (LM1B) and OpenWebText (OWT) datasets. We then distill FLM into a flow map language model (FMLM), whose one-step generation exceeds the 8-step quality of recent few-step discrete diffusion language models. Our work challenges the widely-held hypothesis that discrete noising processes are necessary for generative modeling over discrete modalities and paves the way toward accelerated language modeling at scale.

The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8xA800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to 2.8x and 2.0x in the prefill and decode stages.

Optical flow is an important cue for drone analysis in aerial surveillance, airspace monitoring, and long-range target tracking, where motion remains informative when appearance is weak. Estimating the motion of small Unmanned Aerial Vehicles (UAVs), however, is difficult because distant targets occupy few pixels, have low contrast, and move against complex dynamic backgrounds. Standard optical-flow benchmarks emphasize large, object-centric motions and therefore do not characterize distant aerial targets well. We introduce FlyingDrones, a generated benchmark with dense ground-truth optical flow and segmentation masks derived from physically based rendering, and define resolution-normalized metrics (EPE Norm) that normalize flow by image dimensions. We evaluate 16 state-of-the-art models on small-target motion estimation. Standard RAFT degrades substantially on small targets at low resolution, whereas VideoFlow's multi-frame variant achieves the best drone-region EPE at every tested resolution. Aerial-pretrained two-frame models such as Sea-RAFT and WAFT are competitive at 270$\times$480 but lose their advantage at 1080$\times$1920, and the best fine-tuning domain depends on resolution: KITTI helps when drone motion is sub-pixel-to-few-pixels, while Sintel-style fine-tuning helps once displacements grow with image size. Overall, long-range drone analysis benefits most from multi-frame temporal context, scale-aware evaluation, and resolution-matched fine-tuning.

We present FM-ChangeNet, a pathwise-supervised framework for change detection that reformulates bi-temporal reasoning as continuous transport in feature space rather than static endpoint comparison. Given encoded pre and post-temporal representations, we construct intermediate latent states and learn a time-conditioned velocity field $\hat{v}_\theta(z_t,t)$ along the transformation trajectory. This pathwise formulation constrains the predictor over a continuum of intermediate states, providing a denser and less ambiguous supervision signal than conventional endpoint-only segmentation and enabling the model to capture temporal evolution explicitly. The learned velocity field is not only a transport mechanism but also an interpretable representation of change: its magnitude serves as a spatially localized change cue that helps distinguish true structural variation from nuisance effects such as illumination shifts and spatial misalignment. We develop a hierarchical multi-scale architecture with cross-temporal alignment, time-conditioned coarse-to-fine flow decoding, and a unified objective that couples flow supervision, trajectory consistency, spatial regularization, and segmentation loss. Experiments on remote sensing benchmarks show that the proposed framework produces more structured and robust change representations while achieving state-of-the-art performance.

Second-order optimizers (\eg Shampoo, SOAP) have emerged as compelling alternatives to Adam/AdamW and promise faster per-iteration convergence than AdamW in transformer training. However, they still remain costly in practice: even SOAP still pays substantial optimizer-state memory and eigendecomposition overhead. We ask whether the benefits of SOAP can be retained at nearly the efficiency of AdamW. In this paper, we answer yes with a lightweight yet improved SOAP, Factored One-sided Adam-Moment (FOAM). The path to practical impact is twofold: first, FOAM employs input-only preconditioning, dropping the output-side Kronecker factor to reduce memory and remove one eigendecomposition. A per-row Hessian decomposition identifies this as the right one-sided choice when output curvature is approximately isotropic. Second, FOAM replaces SOAP's dense rotated second moment with a factored Adam-moment estimator based on row-by-column statistics. Under the K-FAC Fisher assumption, this estimator is further proven to be pointwise equivalent to SOAP's update. Our empirical results support the claim that FOAM matches strong SOAP and Muon baselines in validation loss across a diverse scale of large language model (LLM) pre-training, uses optimizer state closer to AdamW, runs within Muon's wall time, and retains robustness across a five-fold learning-rate range. Scaling to 760M parameters, FOAM demonstrates that SOAP's gains need not require SOAP's cost.

LLM benchmarks are usually consumed as scalar leaderboards, where adjacent ranks are read as adjacent capability differences. We argue that this display should be complemented rather than simply trusted or discarded: benchmark scores are useful summaries, but a sorted score table can overstate which model pairs are actually distinguishable. This limitation is especially clear for accuracy- and success-rate leaderboards, whose averages hide task-level structure, and it also appears in pairwise systems such as Elo-style ratings when pairwise evidence is compressed into a total order. We propose LLM tier ranking as a task-level reporting framework for benchmarks. Given benchmark outputs, we construct pairwise wins, losses, and ties between models on each task and aggregate them into ordered tiers using Selective Preference Aggregation (SPA). Cross-tier comparisons are reported only when supported by sufficient task-level agreement; models in the same tier are left unresolved, not declared equal. Across AIME 2025, $\tau$-bench airline, xbench DeepSearch, GDPVal, SWE-Bench Verified, and Chatbot Arena, tier ranking often preserves useful broad capability classes while abstaining from unsupported fine-grained distinctions among frontier models. In Chatbot Arena perturbation studies, tier outputs are more stable than Elo-style total orders, expressing weaker evidence through later tier separation rather than rank reshuffling. We recommend reporting tiered leaderboards, solution paths, top-tier size, singleton-winner thresholds, and pairwise resolution statistics alongside scalar scores.

Existing imitation learning methods enable robots to interact autonomously with the physical environment. However, contact-rich manipulation tasks remain a significant challenge due to complex contact dynamics that demand high-precision force feedback and control. Although recent efforts have attempted to integrate force/torque sensing into policies, how to build a simple yet effective framework that achieves robust generalization under multimodal observations remains an open question. In this paper, we propose ForceFlow, a force-aware reactive framework built upon flow matching. For contact-stage policy design, we investigate force signal fusion mechanisms and adopt an asymmetric multimodal fusion architecture that treats force as a global regulatory signal, combined with a joint prediction paradigm that enhances the policy's understanding of instantaneous force and historical information, thereby achieving deep coupling between force and motion. For task-level hierarchical decomposition, we divide manipulation into a vision-dominant approach stage (VLM-based pointing for target localization) and a touch-dominant interaction stage (force-driven contact execution), with a Vision-to-Force (V2F) handover mechanism that explicitly decouples spatial generalization from contact regulation. Experimental results across six real-world contact-rich tasks demonstrate that ForceFlow achieves a 37\% success rate improvement over the strong baseline ForceVLA while maintaining significantly lower cost. Moreover, ForceFlow exhibits accurate force signal prediction and demonstrates superior performance in contact force self-regulation and zero-shot out-of-distribution (OOD) generalization.

The overarching goal of reinforcement learning (RL) is to efficiently optimize the agent's policy for various environments, tasks, dynamics, and even user preferences, which often consist of multiple and potentially conflicting rewards/objectives. This necessitates the learning of Pareto-optimal policies that can navigate different trade-offs among objectives. However, existing multi-objective RL (MORL) methods, whether trained online or offline, are typically limited to task- or objective-specific policy learning and do not address generalization of learned skills to unseen tasks/objectives. In this work, we propose the Foundation Pareto Flow Policy (FP2) for MORL, inspired by the pre-mid-post-training paradigm underlying recent foundation models. Specifically, we instantiate FP2 as a multi-modal diffusion transformer equipped with a likelihood-free flow matching loss, enabling unified, scalable, and instructable learning of diverse Pareto-optimal trajectories. We further develop a three-stage training pipeline to gradually optimize the policy: (1) pre-training on a large-scale MORL dataset, (2) mid-training on a carefully curated high-quality dataset, and (3) post-training on few collected trajectory data from new tasks and trade-offs. Empirical results demonstrate that FP2 consistently outperforms previous state-of-the-arts on the D4MORL benchmark, achieving superior zero-shot interpolation and extrapolation on complex Pareto manifolds as well as efficient few-shot adaptation to new task/dynamics scenarios.

Automated cardiac MRI quantification is commonly framed either as dense segmentation followed by voxel counting or as direct regression of global measurements. The two paradigms are difficult to unify: segmentation is visually interpretable but label-intensive and output-heavy, while regression is compact but less verifiable. We propose Fourier Contour Learning, which represents cardiac boundaries with compact anatomical Fourier contours. The predicted coefficients define smooth, interpretable contours and enable differentiable volume computation, connecting mask supervision, measurement supervision, and unlabeled-image regularization within one framework. We instantiate this representation with an anatomy-query decoder and cross-slice attention which creates a traceable path from image to contour to clinical measurement. Across three public datasets and an in-house all-phase cohort, our method achieves segmentation accuracy comparable to strong dense-prediction baselines, improves boundary agreement, and reduces clinically relevant volume error while using only a small number of output parameters per structure. We further show that the learned model provides an efficient interface for vision-language question answering and surface reconstruction. These results demonstrate the potential of Fourier contours as a unified intermediate representation for scalable, traceable, and clinically grounded cardiac analysis. Code will be publicly available upon acceptance.

To circumvent the inherent fidelity bottlenecks and optimization misalignment of VAE-based latent diffusion, pixel-space diffusion models have emerged as a compelling end-to-end paradigm. However, existing pixel diffusion models often struggle to balance computational efficiency with the preservation of high-frequency details. They frequently resort to patch-based compression or restricted local decoding, leading to a "spectral compromise" where high-frequency and fine-grained pixel information are suppressed. To address these challenges, we propose \textbf{FrequencyBooster}, a novel framework designed to empower pixel diffusion with full-frequency modeling capabilities without prohibitive overhead. The core of our method is a high-capacity decoder that specializes in extracting exhaustive high-frequency details and low-frequency semantics, the latter of which is derived from a Diffusion Transformer (DiT) backbone. Unlike prior works that sacrifice global context for local refinement, FrequencyBooster leverages high-dimensional feature representations to maintain global structural integrity while achieving superior pixel-level precision. Extensive experiments on ImageNet demonstrate the effectiveness of our approach: our model achieves a state-of-the-art FID of \textbf{1.60} at $256 \times 256$ resolution within only 320 epochs. Furthermore, at $512 \times 512$ resolution, FrequencyBooster attains an FID of \textbf{1.69}, significantly outperforming existing pixel-space and latent-space generative models.

We study online learning in the random-order model, where the multiset of loss functions is chosen adversarially but revealed in a uniformly random order. By extending the *batch-to-online* transformation of Dong and Yoshida (2023), we show that if an offline algorithm enjoys a $(1+\varepsilon)$-approximation guarantee, an *average sensitivity* bound controlled by a function $\varphi(\varepsilon)$, and stability with respect to $\varepsilon$, then we can obtain a *small-loss* regret bound typically of order $\tilde O(\varphi^{\star}(\mathrm{OPT}_T))$, where $\varphi^{\star}$ is the concave conjugate of $\varphi$, $\mathrm{OPT}_T$ is the offline optimum over $T$ rounds, and $\tilde O$ hides polylogarithmic factors in $T$. Our result refines their original $(1+\varepsilon)$-approximate regret guarantee and applies to a broad class of problems, including online $k$-means clustering and online low-rank approximation. We further apply our approach to online submodular function minimization using $(1\pm\varepsilon)$-cut sparsifiers of submodular hypergraphs, obtaining a small-loss regret bound of $\tilde O(n^3 + n^{3/4}\mathrm{OPT}_T^{3/4})$, where $n$ is the ground-set size; we also demonstrate its applicability to online $\ell_1$ regression. Our work sheds light on the power of sparsification and related algorithmic techniques in achieving small-loss regret bounds in the random-order model, without requiring structural assumptions on loss functions, such as linearity or smoothness.

Sign Language Translation (SLT) research has typically focused on pre-segmented clips, a constraint that disconnects models from the continuous reality of real-world communication. To bridge the gap from clips to streams, we formally define \textbf{Streaming SLT}, a realistic task demanding the simultaneous discovery and translation of linguistic events from continuous, untrimmed video inputs. To address this challenge, we propose \textbf{StreamSLST}, the first end-to-end framework for streaming \textbf{S}ign \textbf{L}anguage \textbf{S}egmentation and \textbf{T}ranslation via Deformable Transformer architecture. Unlike cascaded approaches that suffer from error propagation, StreamSLST employs a parallel decoding strategy to jointly optimize \textit{temporal segmentation} and \textit{translation} in a single pass. To enable scalable training on long-form videos, we incorporate a \textit{Sentence-aware Sliding Window Sampler} that crops continuous streams into trainable units. We further design a \textit{Tri-modal Visual-Language Pretraining} stage that aligns pose dynamics with textual semantics before joint optimization. To support comprehensive evaluation, we conduct experiments on two real-world streaming datasets, BOBSL and How2Sign, together with two newly constructed synthetic streaming benchmarks, \textit{Streaming-CSL-Daily} and \textit{Streaming-Phoenix-2014T}. The results demonstrate that StreamSLST consistently outperforms cascaded segmentation-then-translation pipelines. Interestingly, our analysis reveals that learned sentence boundaries can outperform ground-truth boundaries for translation, suggesting that joint optimization drives localization toward semantically informative regions rather than exact temporal endpoints. Our work establishes the first robust benchmark for continuous sign language understanding and paves the way for accessible real-world communication interfaces.

A key problem in the modern study of Deep Learning is predicting and understanding emergent capabilities in models during training. Inspired by methods for studying reactions in quantum chemistry, we present the ``2-datapoint reduced density matrix" (2RDM). We show that this object provides a basis for computationally efficient, unified observables of phase transitions during training. First is the \textit{spectral heat capacity}, which we prove provides early warning signals for learning events. Second is the participation ratio, which reveals the dimensionality of the underlying reorganization associated with a learning event. Remarkably, the top eigenvectors of the 2RDM are directly interpretable, making it straightforward to study the nature of the transitions. We validate across four distinct settings: deep linear networks, induction head formation, grokking, and emergent misalignment. We then discuss directions for future work using the 2RDM.

The automatic synthesis of a program from any form of specification is regarded as a holy grail of computer science. Fueled by LLMs, NL2Code has achieved tremendous success, yet the fundamentally more challenging task of synthesizing programs from input-output behavior, which we refer to as IO2Code, remains largely unsolved. Whereas NL2Code can exploit the semantic alignment between natural language and code acquired during pretraining, IO2Code requires recovering underlying principles from concrete computational behavior, navigating a vast and underspecified hypothesis space. To address this, we propose DIO-Agent, a discovery agent for IO2Code. Our method frames IO2Code as an evolutionary search over discrete program space, in which an LLM serves as the mutation operator and concrete error signals from execution guide each mutation. To prevent the search from wandering into structurally complex yet incorrect dead ends, we introduce the Transformation Priority Premise as a mutation prior that biases the LLM toward the simplest hypothesis consistent with current evidence, progressively escalating from constants to conditionals to iteration only when simpler constructs are insufficient. To facilitate systematic study, we further construct an IO2Code benchmark spanning multiple difficulty levels. Extensive experiments show that DIO-Agent consistently outperforms both traditional program-by-example methods and evolutionary search baselines across all difficulty levels and various LLMs, while substantially surpassing test-time scaling strategies with equivalent sampling budgets. Our code and dataset are available at \url{https://anonymous.4open.science/r/IO2Code}.

Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, yet the internal mechanisms driving these multi-step processes remain opaque. Existing mechanistic interpretability approaches often rely on superficial text-pattern co-occurrences or capture only short-term effects, struggling to trace the long-horizon, outcome-driven influence of internal components such as neurons or sparse features in multi-step reasoning. In this paper, we introduce Integrated Policy Gradient (IPG), a novel, training-free framework that brings policy-based attribution into mechanistic interpretability for LLM reasoning. Grounded in outcome-oriented and sequential-influence-aware principles, it backpropagates outcome-level, non-differentiable behavioral signals (e.g., reasoning outcome correctness) through entire inference trajectories, utilizing path integration to isolate outcome-relevant mediators at component level for reliable attribution. This provides a linear approximation of Causal Mediation Analysis (CMA) in reasoning. Empirical evaluations demonstrate that IPG achieves precise localization and reveals mechanistic insights by identifying sparse, concentrated subsets of internal components associated with specific reasoning behaviors. Our method also enables modulation of reasoning capabilities, offering a powerful and reliable framework for understanding LLM reasoning.

Supervised fine-tuning (SFT) on long teacher trajectories is the dominant method for instilling investigation and reasoning capabilities into open software-engineering (SWE) agents. Under SFT, every retained response is an imitation target, so the student inherits not only the trajectory's outcome but also any flaw in its intermediate steps, including ungrounded leaps and redundant loops. High-quality training data must therefore be jointly effective (each step is grounded and narrows the agent's epistemic gap to the correct fix) and efficient (each step is information-bearing rather than redundant or looping). Existing recipes filter or relabel teacher rollouts using only a binary terminal verifier, which does not directly target these axes and provides no supervision on instances where the teacher fails. Every real issue ships with a developer-authored reference patch $p^\star$ that implicitly testifies to the file paths, runtime behaviors, and conventions a fix presupposes, but the standard pipeline discards it. We propose P2T (Patches-to-Trajectories), which uses $p^\star$ as privileged information during curation, and frames trajectory construction as a bi-objective program over per-step effectiveness and trajectory length. A reverse phase distills $p^\star$ into a latent process graph $G^\star$ of contextual facts and solution milestones, encoding dense intermediate anchors in constructive ordering. A forward phase curate trajectories from blinded teacher continuations, scoring per-step progress against $G^\star$ under a leakage-blocking groundedness check and committing the shortest segments that retain effectiveness. Using only 1.8k curated SWE-Gym instances, P2T improves both axes simultaneously over outcome-filtered SFT and its tool-error-masking variant: on SWE-bench Verified, it lifts Pass@1 by up to +10.8 points while cutting per-instance inference cost by 15%, with consistent gains on SWE-bench Lite and across two teachers. A size-matched ablation and qualitative analysis further isolate per-trajectory quality from data scale.

Reinforcement learning from verifiable rewards (RLVR) has shown strong promise for LLM reasoning, but typical outcome-based RLVR methods remain inefficient on hard problems. Correct final-answer rollouts are rare, and standard sample-level credit assignment fails to leverage partial reasoning progress embedded in unsuccessful attempts. To address this problem, we introduce **SCRL** (**S**ubproblem **C**urriculum **R**einforcement **L**earning), a curriculum reinforcement learning framework built on verifiable subproblems derived from reasoning chains. Given a reference solution, SCRL derives a series of verifiable subproblems and constructs a subproblem curriculum, with the final subproblem fixed as the original problem. This converts partial progress on hard problems into verifiable learning signals. Algorithmically, we propose *subproblem-level normalization*, a training technique based on RLVR that normalizes rewards independently at each subproblem position within the rollout group. By assigning the resulting advantages to the corresponding answer spans, we enable finer-grained credit assignment without external rubrics or reward models. Our theoretical analysis shows that this subproblem curriculum makes hard problems more learnable by lifting them out of gradient dead zones, with larger relative gains as the original problem becomes harder. Across seven mathematical reasoning benchmarks, SCRL outperforms strong curriculum-learning baselines, yielding +4.1 and +1.9 average-point gains compared to GRPO on Qwen3-4B-Base and Qwen3-14B-Base respectively. On three hard benchmarks (AIME24, AIME25, and IMO-Bench), SCRL further yields point gains of +3.7 in pass@$1$ and +4.6 in pass@$64$ on Qwen3-4B-Base, suggesting improved exploration on hard reasoning problems. Further ablations show that the proposed credit assignment is effective and that the gains do not require highly curated subproblems or strong external generators, highlighting SCRL as a practical curriculum framework that enables fine-grained credit assignment for LLM reasoning.

Multimodal Large Language Models (MLLMs) can listen and see, but how do audio and visual signals actually travel through the network to shape an answer? Despite their growing role in research and real-world applications, the internal pathways through which audio and visual tokens influence the final prediction remain poorly understood. In this study, we examine audio-visual information flow inside Audio-Visual Large Language Models (AVLLMs), tracing how AVLLMs route, utilize, and integrate audio and visual information across two input configurations, audio-visual video and multiple interleaved audio-visual items. We find that for audio-visual video, AVLLMs follow the sequential information flow pathway established for VLMs and VideoLLMs, with audio and visual contribution flowing along this pathway in proportion to the task's reliance on each modality. In settings with multiple interleaved audio-visual items, this routing shifts to different parallel streams. Furthermore, we demonstrate that audio-visual and other token types can be discarded once their information is transferred to LLM, with minimal impact on the model's prediction or even slight improvement, generalizing across multiple tasks and datasets, enabling more efficient inference. These findings hold across multiple models and scales, Qwen2.5-Omni and Video-SALMONN2 Plus at 3B and 7B scales, leading to hypotheses on why these flow structures emerge. Together, these results deliver the first coherent picture of how AVLLMs orchestrate sound and sight inside the network and lay the groundwork for the next wave of interpretability, design, and efficiency advances in audio-visual and broader MLLMs.

Graph domain adaptation (GDA) aims to transfer knowledge from a labeled source graph to an unlabeled target graph. Existing GDA methods typically assume that the source graph provides sufficiently complete information for reliable cross-domain alignment. However, this assumption is often unrealistic in practice, where only a small sampled source graph may be available due to privacy, annotation, storage, or data-collection constraints. In this paper, we study Cross-Scale Graph Domain Adaptation (CSGDA), where adaptation is performed from a limited-scale source graph to a larger target graph. Our empirical results show that CSGDA exhibits larger local and global discrepancies than conventional GDA, as the source graph provides only limited local structures and limited coverage of the global structure. To address this challenge, we propose a local-global structural alignment framework that, at the local level, augments neighborhood structures to compensate for limited local structural information. At the global level, it extracts global structural patterns and aligns source and target distributions. An adaptive GCN-MLP module then integrates the compensated local structures and extracted global patterns for domain alignment and target prediction. Experiments across multiple graph adaptation benchmarks and sampling ratios demonstrate that our method consistently improves adaptation in the presence of severe source-graph incompleteness.

Knowledge distillation from large reasoning models (LRMs) often uses raw teacher trajectories as supervision. However, this practice creates a mismatch between the goal of distillation and the nature of teacher trajectories. Advanced LRMs typically improve their reasoning capabilities through reward-driven post-training, such as RLHF. This training optimizes them as strong solvers, rather than as teachers that produce reasoning processes smaller models can easily learn. Since smaller models have different "cognitive capacities" compared to their larger teachers, directly imitating raw teacher trajectories can sometimes be ineffective and often requires a substantial number of high-quality samples. To bridge this mismatch, we propose Cognition-Aligned Reasoning Distillation (CARD), an RL-based framework that transforms solver trajectories into teaching trajectories aligned with the student’s cognition. Avoiding expensive resampling, our framework directly transforms existing trajectories. We optimize CARD with two reward signals: (1) a Distribution Alignment Reward that matches the student’s predictive distribution, and (2) a Logical Coherence Reward that reduces misleading failed reasoning steps. Comprehensive evaluations across multiple student models and settings demonstrate that CARD outperforms distillation from raw teacher trajectories in in-domain settings while achieving substantial gains in out-of-distribution performance.

Direct Preference Optimization (DPO) is widely used to reduce hallucination in Multimodal Large Language Models (MLLMs). Prior studies and our observations show that during DPO training, the log probabilities of chosen and rejected responses can decrease simultaneously, revealing the squeezing effect. For MLLMs, the redistributed probability mass may move toward responses driven by language priors rather than visual evidence, weakening the role of the image and amplifying hallucination. Existing squeezing-aware methods constrain rejected updates with $\mathcal{V}$-usable information or generation confidence, while methods with anchor terms regularize the DPO implicit reward of chosen responses. Neither explicitly grounds the control signal in visual evidence. We propose Visual Guided Direct Preference Optimization (VGDPO), which relaxes negative updates on rejected responses according to multimodal squeezing risk. VGDPO estimates this risk by combining token-level visual dependency with phrase-level hallucination localization, defining a hallucination ratio for each rejected response. We further apply a related reweighting principle to a visual contrastive objective, providing additional supervision for chosen responses under the original image. Experiments on four hallucination benchmarks and three MLLM backbones show that VGDPO improves DPO log probability dynamics and reduces multimodal hallucination while maintaining informative responses.

Long-context inference in large language models is bottlenecked by the quadratic cost of full attention. Existing efficient alternatives often rely either on native sparse training or on heuristic token eviction, creating an undesirable trade-off among efficiency, training cost, and accuracy. In this work, we show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation. Our approach is built on three observations: (1) only a small subset of attention heads truly requires full long-context processing; (2) long-range retrieval is governed primarily by a low-dimensional subspace, allowing relevant tokens to be retrieved efficiently with a 16-dimensional indexer; and (3) the useful token budget is strongly query-dependent, making dynamic top-$p$ selection more suitable than fixed top-$k$ sparsification. Based on these insights, we propose RTPurbo, which retains the full KV cache only for retrieval heads and introduces a lightweight token indexer for sparse attention. By exploiting the model's intrinsic sparsity, RTPurbo achieves sparsification with only a few hundred training steps. Experiments on long-context benchmarks and reasoning tasks show that RTPurbo preserves near-lossless accuracy while delivering substantial efficiency gains, including up to a 9.36$\times$ prefill speedup at 1M context and about a 2.01$\times$ decode speedup. These results suggest that strong sparse inference can be obtained from standard full-attention training without expensive native sparse pretraining.

Multi-modal semantic segmentation benefits from complementary sensors, but this benefit is fragile when some modalities are degraded or missing at test time. Many existing methods rely on dedicated fusion modules or multi-branch cross-modal encoders, and the fusion often becomes unreliable when modalities are absent, leading to substantial performance drops. We propose FuseAdapt, a parameter-efficient framework that performs fusion within the adaptation space of a frozen vision foundation model (VFM), exploring whether VFM priors can serve as an effective shared backbone for multi-modal segmentation. Instead of attaching a standalone fusion network, FuseAdapt assigns each modality a lightweight adaptation branch and performs adaptation-space fusion by routing trainable adaptation updates across modalities inside the frozen backbone. The pretrained VFM remains fixed while modality-specific corrections and explicit cross-modal interaction are learned through the adaptation paths. To handle missing modalities, we further introduce Predictive Modality Modeling (PMM), a direct representation-level objective for cross-modal redundancy. During training, PMM masks a target modality and predicts its embedding from the visible ones, supervised by a teacher target computed under full-modality input. PMM is used only during training and adds no inference-time overhead. Across three multi-modal segmentation benchmarks (MCubeS, DELIVER, MUSES), FuseAdapt consistently improves both modality-complete and modality-incomplete segmentation with a small trainable budget, yielding up to +3.01 mIoU under full modalities and +12.56 mIoU under missing modalities over prior state-of-the-art.

Multimodal Large Language Models (MLLMs) are typically aligned with caption-pair data, but not all answer-token supervision is equally visual. In LLaVA-style alignment training, many tokens are predictable from language context alone, yet their losses can still update vision-related parameters. Such language-dominant updates may add noisy supervision for tokens that require visual evidence. To diagnose this vulnerability, we probe the model against visual counterfactuals to reveal the intrinsic grounding requirements of each token, distilling them into two soft indicators: visual necessity and evidence specificity. Guided by this diagnostic insight, we present $\textbf{GEAR-Align}$, a controlled post-training framework that makes supervision destination explicit. During training, it converts token-level visual reliance into parameter-level gradient allocation: all answer tokens update language-side parameters, high-necessity tokens update route-flow parameters, and the high-specificity subset additionally updates visual-evidence parameters. Across two controlled Qwen3-backbone settings, GEAR-Align improves hallucination-sensitive and local-evidence-heavy benchmarks while remaining competitive on general multimodal benchmarks. These results provide evidence for token-specific visual gradients as a useful principle for controlled MLLM post-training.

Accurate characterization of deep subsurface structures is a fundamental problem in geoscience, but geophysical inversion is inherently non-unique and therefore requires joint interpretation of gravity, magnetic, electromagnetic, and seismic observations. Despite recent advances in deep learning for geophysical inversion and PDE-based scientific modeling, existing large-scale open datasets are largely limited to a single modality and rarely provide physically consistent multi-source responses generated from the same 3D geological target. This limitation hinders the development and evaluation of multimodal joint inversion methods for deep subsurface characterization. We introduce GEMS-3D, a large-scale synthetic benchmark for 3D geological multiphysics learning. GEMS-3D couples four forward modeling engines—$\textbf{g}$ravity, $\textbf{e}$lectromagnetic, $\textbf{m}$agnetic, and 3D acoustic $\textbf{s}$eismic—to generate sample-aligned observations from shared geological realizations. Guided by facies-aware rock-physics relationships, the dataset provides spatially aligned volumes of P-wave velocity ($v_p$), density ($\rho$), resistivity ($\phi$), and magnetic susceptibility ($\chi$) across nine representative classes of complex geological anomalies, including dike swarms, salt domes, and gas reservoirs. Each sample includes gravity responses, magnetic responses, electromagnetic responses, and 3D acoustic seismic data, together with acquisition metadata and anomaly labels. By packaging co-registered property volumes, anomaly support, acquisition metadata, and aligned multiphysics responses as sample-level bundles, GEMS-3D provides a unified benchmark for operator learning, multimodal inverse interpretation, and missing-modality robustness in deep subsurface settings. The anonymized forward-modeling framework and corresponding dataset link is available at https://anonymous.4open.science/r/GEMS-3D-E4CA.

Learning the dynamics of interacting physical systems is a central challenge in machine learning for the physical sciences. Prevailing architectures such as Graph Neural Networks (GNNs) and feedforward models often fail to explicitly capture the compositional nature of physical laws. Instead of modeling the forces that govern a system, these networks learn to approximate aggregate state transitions. As a result, they lack a causal understanding of the underlying physics interactions, leading to poor generalization in out-of-distribution systems with varying numbers of entities or novel configurations. In this work, we introduce Compositional Potential Minimization (CPM), a framework that casts simulation as an energy minimization procedure over a composition of learned force potentials. Each force component is represented as a separate, interpretable energy landscape, and interaction dynamics emerge from their addition. Inference in CPM is framed as a global energy minimization problem over an arbitrary number of composed functions. Because CPM learns to disentangle separate forces during training, it enables seamless generalization during test time to larger-scale systems and unseen configurations through straightforward compositions of the learned energy components. Our experiments show that CPM significantly outperforms previous state-of-the-art GNN-based solutions in 2D, 3D, and particle-based simulations.

Influence functions (IFs) approximate the effect of removing or reweighting training data on a trained model, but their linear approximation often lacks the accuracy required for reliable post hoc model editing. We identify two culprits and propose generalized influence functions (GIFs) and pseudo-LiSSA (p-LiSSA) to address each in turn. First, classical IFs update all parameters indiscriminately, leading to unnecessary updates to data-irrelevant parameters. GIFs instead restrict edits to a \textit{data-relevant subspace} and leave the remaining parameters fixed. We theoretically show that GIFs achieve a tighter error bound than classical IFs, and that proper subspace selection requires avoiding low-curvature directions while maintaining strong alignment with the target gradient. Second, LiSSA iterations are destabilized by negative Hessian eigenvalues, which existing methods handle via $\ell_2$ damping---at the cost of a biased inverse-Hessian-vector product (iHVP) estimate. p-LiSSA eliminates this bias by solving the \textit{restricted inverse-Hessian problem} without such damping. We further show that the presence of low-curvature directions slows p-LiSSA convergence, and that the data-relevant subspace selected by GIFs mitigates this. Across five image and text unlearning benchmarks, GIFs update only 10% of the parameters yet preserve accuracy up to 2.12 percentage points better than IF-based baselines, while improving unlearning performance by up to 2.73 percentage points. Moreover, p-LiSSA improves restricted iHVP estimation by reducing the residual by 9.6$\times$ over existing iHVP baselines, while using 19.8% less memory with comparable computation time.

Despite rapid advances in text-to-image generation, faithfully realizing user intent remains challenging, often requiring manual multi-turn trial and error. To automate this process, existing systems rely on either simple prompt rewriting or closed-loop agents driven by hand-crafted rules, rather than learning to adapt actions to the evolving generation process. In this paper, we reformulate image generation as a state-conditioned action-making problem and propose Generation Navigator, a multi-turn T2I agent that learns to dynamically steer the generation trajectory and output the next action. However, training this agent via reinforcement learning introduces a critical credit assignment challenge: naively rewarding a trajectory based solely on a single state assigns equal credit to all actions in the rollout, ignores the quality dynamics across turns, and fails to distinguish actions that improve the trajectory from those that degrade it or waste turns without progress. We resolve this with PRE-GRPO (Peak-Retention-Efficiency Group Relative Policy Optimization), a trajectory-level reinforcement learning objective that explicitly rewards discovering a high-quality image (Peak), avoiding subsequent quality degradation across turns (Retention), and minimizing unnecessary turns (Efficiency). Experiments show substantial improvements across benchmarks, reaching a WISE score of 0.90 and 79.06% reasoning accuracy on T2I-ReasonBench.

Expressive generative policies such as diffusion and flow models are appealing for maximum-entropy (MaxEnt) online reinforcement learning because of their ability to model multimodal and highly non-Gaussian action distributions. However, training effective soft generative policies faces two obstacles that often arise together. First, marginal action densities are often unavailable, so existing methods typically rely on entropy bounds, heuristic proxies or approximations. Second, iterative shared-parameter samplers raise inference cost and require backpropagation through time over repeated network evaluations, increasing memory cost and destabilizing policy optimization. These obstacles motivate us to seek a generative policy that exposes a tractable MaxEnt objective while requiring only a single sampled actor forward pass for action generation. To this end, we propose soft generative actor-critic (SoftGAC), whose actor defines a stochastic bridge from a fixed base latent to a terminal action latent in pre-tanh space. This structured bridge allows us to lift the MaxEnt objective as an analytically tractable path-wise relative-entropy objective against a high-entropy reference process. In practical finite-step implementation, this relative entropy reduces exactly to sampled transition control energy and thus provides principled soft regularization. Moreover, we keep the single-pass actor lightweight by using small step-specific bridge transitions, each evaluated only once per sampled action, while maintaining a parameter budget comparable to strong actor baselines. Extensive experiments on challenging continuous-control benchmarks show that \myalgo attains higher or competitive returns than strong generative policy baselines, including diffusion and flow-matching policies, while staying in the low-latency regime of one-pass actors and avoiding the much higher cost of many-step diffusion samplers, showing considerable improvements in the compute-return tradeoff.

Existing generative models for earth observation (EO) predominantly rely on fine-tuning natural image priors, which limits their scalability and introduces perspective biases that conflict with geospatial constraints. To address this, we introduce GeoCore-9B, a 9-billion-parameter generative foundation model, which is the first of its scale to be trained from scratch exclusively on EO data. Unlike previous EO foundation models, GeoCore-9B is built upon a Flow Matching-based Diffusion Transformer (DiT) and natively conditions generation on text descriptions and continuous geospatial metadata, including ground sample distances, latitudes, and longitudes. To overcome the convergence and spatial disorientation challenges of training at this scale, we propose a Geospatial Semantic Alignment loss. This objective distills structural Earth surface priors (e.g., terrain and urban areas) from a frozen specialist teacher network, constraining the diffusion latent trajectory during training without adding inference overhead. Pre-trained on the global-scale Git-10M dataset, GeoCore-9B demonstrates strong downstream versatility. Beyond standard proxy generative tasks, we show that GeoCore-9B can be effectively adapted for practical EO applications, including highly challenging tasks such as cloud removal and SAR-to-optical cross-modal translation. Extensive evaluations confirm that GeoCore-9B establishes new state-of-the-art performance in both visual fidelity and geographic structural accuracy. Code and pretrained checkpoints will be released upon acceptance.

Flow matching learns a velocity field whose ordinary differential equation transports a simple reference distribution to a data distribution. Existing theory typically controls this velocity through ambient, isotropic Lipschitz regularity, which becomes extremely pessimistic when data concentrate near a low-dimensional manifold. In this regime, standard error bounds can blow up drastically near the terminal time and fail to explain why flow matching should remain stable. We study flow matching for distributions obtained by smoothing a density on a compact manifold with small Gaussian noise. Our first result is a tube-local Lipschitz regularity theorem showing that, on the relevant neighborhood of the manifold, the population flow-matching velocity has a Jacobian bound that grows only at the inverse effective noise scale. This improves substantially over the stronger singular behavior suggested by existing ambient analyses. Our second result gives a directional description of the velocity Jacobian: tangential directions exhibit mild growth, normal directions are strongly contractive, and tangential-normal coupling is lower order. This reveals that the apparent singularity is highly structured rather than uniformly harmful. These results provide geometric foundations for sharper stability guarantees and suggest a path toward intrinsic-dimension generalization bounds for flow matching on manifold-concentrated data.

Zeroth-order optimization (ZOO) is a promising approach for memory-efficient fine-tuning of Large Language Models (LLMs). However, scaling ZOO to larger models on a single GPU necessitates aggressive quantization, which often destabilizes training. A primary cause of this instability is that existing methods treat strictly positive quantization scale parameters as Euclidean variables, potentially leading to boundary violations and optimization difficulties. To address this issue, we propose Hyper-Octant Zeroth-order Optimization (HoZO), a geometry-aware framework that formulates fine-tuning as Riemannian optimization on the positive orthant manifold. By constructing a novel geodesically complete metric and deriving the closed-form exponential map, HoZO inherently enforces positivity constraints without bias-inducing projections. On the theoretical side, we prove that HoZO achieves the optimal oracle complexity for zeroth-order methods. On the empirical side, HoZO outperforms baselines across six downstream tasks and three model sizes. Notably, it achieves a $34\times$ memory reduction compared to first-order fine-tuning, enabling the stable fine-tuning of a 4-bit quantized Llama-2-70b model on a single 48GB GPU.

Kolmogorov–Arnold Networks (KANs) replace fixed activations in deep architectures with learnable univariate edge functions, making the choice of edge parametrisation central. Existing variants rely on fixed bases such as splines, polynomials, or Fourier features, which impose a function-space geometry before data are observed. We introduce geometry-constrained KANs, a family of edge activations derived from Banach duality maps in which the geometry itself is learned through a scalar exponent $p > 1$ per edge. This exponent controls the qualitative response: sub-Euclidean values produce sharp threshold-like behaviour, $p = 2$ recovers the linear regime, and larger values produce flatter responses near the origin. Beyond expressivity, the exponent also modulates sensitivity: smaller values reduce amplification of perturbations, providing an implicit regularisation effect without introducing an explicit shrinkage hyperparameter. Across 50 Feynman symbolic regression equations, geometry-constrained KANs match strong fixed-basis baselines on clean data while substantially improving robustness under distribution shift. In particular, they outperform cross-validated spline baselines in 85 of 90 data-scarcity settings and degrade significantly less under noise (3.7× versus 21.6× for splines). Learned exponents are interpretable and stable, revealing consistent geometric structure across equation families and input dimensions.

Packed high-token 3D fields are not just long-sequence modeling: once a volume is packed into tokens, sequence order is no longer the geometry of the sample. This creates a general failure mode for attention over packed geometric data: sequence-index positional mechanisms can learn the memory layout rather than the object. We introduce Pack3D, a geometry-aware local--global attention operator that separates memory layout from attention geometry. Pack3D keeps local neighborhood attention exact, compresses only distant context into contiguous 3D block summaries, and indexes attention with axial RoPE on true 3D token-grid coordinates rather than packed sequence positions. With RoPE applied before block compression, each query interacts with a geometry-aware mixture of block tokens, so compressed global context remains spatially interpretable after packing. A synthetic repacking benchmark makes the failure mode explicit: learned positional embeddings and packed-sequence RoPE fail under layout changes, whereas axial grid-coordinate RoPE remains invariant. On Allen WTC--11 hiPSC microscopy, our main claim-grade domain, Pack3D improves matched learned-position local--global attention by 0.0978 best macro Dice and 0.1248 final macro Dice at 262k tokens and depth 8. In an Allen-15 joint comparison, Pack3D also improves over a parameter-matched TokenGrid 3D U-Net on both best and final macro Dice. Inference-time ablations show that the trained model genuinely uses its compressed summaries, and the fused forward/backward implementation is $6.49\times$ lower latency than dense SDPA at 65k tokens.

Recent advances in robotic manipulation have made rapid progress in mapping visual observations and language instructions to actions. However, reliable manipulation in open-world environments requires robust spatial understanding, the ability to identify targets from complex layouts (spatial reasoning) and adapt to scenes that continuously evolve through interaction (spatial memory). Yet existing embodied benchmarks fall short in evaluating spatial understanding: they often introduce strong visual cues that bypass spatial reasoning, and focus on static scenes that offer little insight into an agent’s capacity for spatial memory. To address this gap, we introduce GeoMind, a systematic benchmark for structured spatial understanding through tabletop manipulation. GeoMind comprises 54 evaluated task variants derived from 27 canonical task designs organized around spatial reasoning and spatial memory. It also incorporates a scalable pipeline that automates the generation and annotation process of task instances with verified targets and executable interactions. Extensive evaluations across diverse embodied agents reveal their consistent weaknesses in spatial understanding. Our findings highlight that these agents struggle to track identities after spatial changes and convert relational or geometric constraints into executable actions.

Volumetric radar nowcasting asks a model to forecast where storm structures move and how reflectivity evolves through height. The recent \stcgs{} method makes this setting practical by tracking 3D radar volumes with spatiotemporally coherent Gaussian parcels, but its forecasting head is designed to evolve all future Gaussian attributes with a unified deterministic regressor. This is a poor match to convective evolution: geometry is dominated by coherent transport, while strong-echo radiometry is local, intermittent, and uncertain. We present GeoRad-3D, a factorized forecasting head for coherent Gaussian radar states. The key idea is simple: move geometry first, predict the stable radiometric continuation next, model only the residual uncertainty, and form the final nowcast after rendering. For geometry, a height-conditioned shared field captures volume-level advection and mild coherent deformation, while a parcel-wise correction restores local parcel drift, shape change, and cross-height motion. For radiometry, a deterministic base predictor outputs the stable continuation, and conditional flow matching samples residuals around this base to model the local uncertainty concentrated around strong echoes. At inference, multiple aligned futures are rendered and aggregated through a base-anchored quantile summary in radar-product space, where threshold skill is measured. Under the published NEXRAD protocol, GeoRad-3D improves Pool4 CSI over the strongest published baseline by about 45\%, 82\%, and 124\% at 20, 30, and 40 dBZ. It further reduces LPIPS-Radar by 31.0\%, indicating better perceptual fidelity of rendered storm structures.

Multimodal Large Language Models (MLLMs) have significantly advanced instruction-driven perception through the ``Embedding-as-Mask'' paradigm. However, extending this capability to Remote Sensing (RS) reasoning segmentation remains exceptionally challenging under cluttered geographical contexts and extreme scale variations, often leading to \textit{Spatial Fidelity Degradation} and \textit{Target Misidentification}. We identify a critical yet underexplored bottleneck: the sequential alienation of the segmentation token \texttt{[SEG]} from its initial visual grounding within the autoregressive generation paradigm. Produced only at the end of long reasoning chains, \texttt{[SEG]} relies on visual evidence that has been progressively diluted by linguistic abstraction and autoregressive noise. This information decay is particularly catastrophic for RS targets with minuscule pixel footprints, where even marginal fidelity loss leads to localization failure. To address these bottlenecks, we propose \textbf{GeoReason}, a novel framework designed to fortify logical deduction and preserve spatial integrity. First, we introduce an \emph{Anticipatory Prior} mechanism that injects the \texttt{[SEG]} token directly into the initial user query, shifting it from a terminal output to a primary condition. This paradigm shift transforms the token into a persistent spatial anchor, preventing information decay during long-form linguistic inference. Second, we enhance visual reasoning via a \emph{Location-aware Chain-of-Thought (CoT)} that enforces coarse-to-fine localization using bounding boxes, coupled with a \emph{Latent Patch-to-Token Alignment (PTA) Loss} that explicitly aligns the latent segmentation token with physical RS textures. Extensive experiments demonstrate that GeoReason consistently outperforms state-of-the-art methods on the Earthreason and RRSIS-D benchmarks.

In this paper, we present gfnx, a fast and scalable package for training and evaluating Generative Flow Networks (GFlowNets) written in JAX. gfnx provides an extensive set of environments and metrics for benchmarking, accompanied with single-file implementations of core objectives for training GFlowNets. We include synthetic hypergrids, multiple sequence generation environments with various editing regimes and particular reward designs for molecular generation, phylogenetic tree construction, Bayesian structure learning, and sampling from the Ising model energy. Across different tasks, gfnx achieves significant wall-clock speedups compared to Pytorch-based benchmarks (such as torchgfn library) and author implementations. For example, gfnx achieves up to 55 times speedup on CPU-based sequence generation environments, and up to 80 times speedup with the GPU-based Bayesian network structure learning setup. Our package provides a diverse set of benchmarks and aims to standardize empirical evaluation and accelerate research and applications of GFlowNets.

Generating accurate radiology reports from 3D medical volumes remains prone to hallucination. Existing grounding strategies often depend on external visual foundation models or on fixed medical phrase queues. The former passes segmenter-derived evidence to generation, where segmentation errors are hard to suppress. The latter has limited coverage of long-tail and compositional abnormalities. We therefore propose HeatRAD, a glance-before-tell framework that first routes suspicious slice evidence into an organ-ordered visual prefix, then generates reports from this localized evidence without explicit grounding supervision. HeatRAD uses a slice-centric signed heat-conduction encoder that performs efficient spectral mixing while separating smooth anatomy from lesion-like residues. \method{} then scores suspicious slices, binds their evidence to coarse organ contexts, and assembles an organ-ordered visual prefix, guiding the decoder toward localized abnormalities instead of redundant volumetric context. Experiments on three 3D radiology benchmarks show consistent gains in clinical fidelity and hallucination-related metrics, supporting anomaly-first evidence selection as an annotation-efficient paradigm for 3D report generation.

Diffusion large language models (LLMs) have recently emerged as a promising alternative to autoregressive LLMs, enabling parallel generation through iterative denoising. Recent guidance methods improve diffusion LLMs by introducing a weak model to guide the prediction of the base model during denoising. However, the influence of the weak model during denoising remains less clear, and computing its prediction requires an additional forward pass at each denoising step. We present in this paper an in-depth analysis of how diffusion LLMs use contextual information during denoising. Our analysis reveals that diffusion LLMs rely more heavily on local context while often underutilizing global context. Based on this observation, we introduce Global Context Guidance (GCG), a feature-space guidance method that enhances global-context utilization without constructing an explicit weak model. Extensive experiments on LLaDA-8B, Dream-7B, and LLaDA-1.5 across diverse benchmarks demonstrate that GCG consistently outperforms existing guidance methods, while incurring much lower inference overhead. We will release our code upon acceptance.

To date, there exist almost no culturally-specific evaluation benchmarks for large language models (LLMs) that cover a large number of languages and cultures. In this paper, we present Global PIQA, a participatory commonsense reasoning benchmark for over 100 languages, constructed by hand by over 350 researchers from over 65 countries around the world. The 141 language varieties in Global PIQA cover five continents, 19 language families, and 24 writing systems. In the non-parallel split of Global PIQA, over 50% of examples reference local foods, customs, traditions, or other culturally-specific elements. In the parallel split, we translate more "culturally agnostic" commonsense reasoning questions into 131 language varieties, for direct cross-lingual comparisons. In both splits, all examples have been verified by native speakers of the languages. We find that state-of-the-art LLMs perform well on Global PIQA in aggregate, but they exhibit weaker performance in lower-resource languages (e.g. up to a 68% accuracy gap between languages in the parallel split). Global PIQA highlights that in many languages and cultures, everyday knowledge remains an area for improvement in LLMs, alongside more widely-discussed capabilities such as complex reasoning and expert knowledge. Beyond its uses for LLM evaluation, Global PIQA provides a glimpse into the wide diversity of cultures in which human language is embedded.

Finding frequently occurring subgraph patterns or network motifs in neural architectures is crucial for optimizing efficiency, accelerating design, and uncovering structural insights. However, as the subgraph size increases, enumeration-based methods are perfectly accurate but computationally prohibitive, while sampling-based methods are computationally tractable but suffer from a severe decline in discovery capability. To address these challenges, this paper proposes GraDE, a diffusion-guided search framework that ensures both computational feasibility and discovery capability. The key innovation is the $\underline{\textnormal{Gra}}\textnormal{ph}$ $\underline{\textnormal{D}}\textnormal{iffusion}$ $\underline{\textnormal{E}}\textnormal{stimator}$ (GraDE), which is the first to introduce graph diffusion models to identify frequent subgraphs by scoring their typicality within the learned distribution. Comprehensive experiments demonstrate that the estimator achieves superior ranking accuracy, with up to 114\% improvement compared to sampling-based baselines. Benefiting from this, the proposed framework successfully discovers large-scale frequent patterns, achieving up to 30$\times$ higher median frequency than sampling-based methods.

We propose a novel solution to reward hacking in reinforcement learning. Our setup involves a misspecified reward function corresponding to task completion, and a classifier that flags invalid solutions. As in frontier LLM development, the classifier makes systematic errors. Consequently, reward penalties and data filtering based on the classifier fail to prevent reward hacking. To address this limitation, we introduce GRAFT: Gradient Routing Adapter Fine Tuning. Rather than penalizing or filtering flagged behaviors during training, we constrain (``route'') gradient updates to specific parameters. This has two benefits. First, the model does not learn to exploit classifier errors, i.e., it does not learn to hide the unintended behavior. Second, the unintended behavior can be disabled by ablating the parameters to which they were localized. We demonstrate these benefits in a variety of RL environments where reward penalties result in monitor-subverting policies. Additionally, we show that our method remains effective when we scale model size and task complexity by distilling frontier LLM trajectories into Qwen3-32B. Overall, we show that GRAFT is a promising approach for leveraging unreliable monitors during post-training in environments where true performance is difficult to evaluate.

Predicting labels for graph-structured data is crucial in many scientific applications. Recently, Gaussian processes (GPs) with graph-level inputs have been proposed as flexible, non-parametric models for classification tasks. In this work, we extend this framework to regression tasks and simplicial complexes (SCs), enabling edge-level attributes and attributes supported on higher-order simplices. Drawing on the rich literature on Hodge theory in machine learning, we enhance the resulting SC representations via the Hodge decomposition, capturing homological information such as holes. We introduce Hodgelet representations as rich, learnable topological descriptors of simplicial complexes and show that our framework improves predictions across various applications. This paves the way for broader use of GPs in graph-level and SC-level prediction tasks.

We introduce **Graph Cascades**, a mesoscopic rewiring strategy for Graph Neural Networks (GNNs) and Graph Transformers (GTs) that captures intermediate-scale graph structure beyond purely local edges or fully global attention. Using contagion-based diffusion processes, Graph Cascades constructs, in $\mathcal{O}(|V|+|E|)$ time, an auxiliary graph where node pairs supported by repeated multi-hop reinforcement are promoted to direct neighbors. We theoretically characterize when reinforcement-based rewiring helps: sufficient conditions under which a reinforcement-based edge-selection rule has higher label agreement than direct adjacency, an SBM witness in which two-hop reinforcement is perfectly homophilic, and formalizing mesoscopic connectivity via graph effective resistance. Empirically, across node-classification benchmarks, Graph Cascades improves multiple GNN and sparse-GT backbones, with the most reliable gains observed on heterophilic and moderate- to high-degree homophilic graphs. The theoretical conditions also identify regimes where mesoscopic rewiring is unlikely to be beneficial -- low-degree regular graphs and graphs with structural bottlenecks -- and these predictions match the observed failures. We additionally observe tight correlations between performance and structural properties in the rewired graphs.

Prioritized Sweeping (PS) accelerates model-based reinforcement learning by selecting backups according to residual magnitude. In nonstationary reward settings, however, the canonical priority score is shortsighted: after a localized reward shift, residuals propagate only through realized backups, so bottlenecked or topologically distant state estimates may remain static under a limited replanning budget. We introduce Graph Topology Augmentation for Prioritized Sweeping (GTA-PS), a concrete instance of a broader topology-augmentation principle for priority-based planning. GTA-PS constructs a policy-induced transition graph and augments the residual key with a mixing of regularized directed-Laplacian potentials that diffuses residual information through forward and backward graph structure. The topology weight is controlled by a scheduler based on the Second Largest Eigenvalue Modulus (SLEM), allowing the queue to adapt to the chain's mixing regime. We prove that the forward potential coincides with discounted residual propagation at a canonical regularization parameter and show that GTA-PS gives nonzero priority to states that standard PS can leave blocked after sparse reward shifts. Tabular experiments on FourRooms and GARNET domains demonstrate improved replanning efficiency over standard PS under both exact DP and Dyna-style host planners.

Models that recognize when their predictions are unreliable inspire greater trust than those that make confident but incorrect decisions. While recent advances enable uncertainty quantification in GNNs, understanding why a model is uncertain, whether due to noisy features, conflicting neighborhoods, or sparse training coverage, remains unsolved. We address this gap with GraphUAT, the first uncertainty attribution framework for GNNs, which identifies the sources of uncertainty rather than aiming to improve its quantification. GraphUAT employs a teacher-student paradigm, where the trained GNN's behavior is distilled into an uncertainty-aware student that disentangles aleatoric and epistemic uncertainties. We then use targeted probes to trace the student's uncertainty back to the responsible features and structural components. Experiments across benchmarks show that removing the identified elements consistently reduces the teacher's uncertainty, validating the faithfulness of our attributions.

3D scene graphs provide a structured representation of complex environments by encoding objects, their semantic attributes, and the spatial and functional relationships between them. Current approaches for 3D scene graph generation suffer from several fundamental limitations. They rely on complex multi-stage pipelines with explicit intermediate representations, making systems fragile and prone to error propagation. They assume access to ground-truth object annotations during inference, which deviates from real-world scenarios. They depend on proprietary models, hindering open-source deployment, or incur prohibitively slow inference. We present GraphWrit3R, a simple end-to-end method that takes a 3D point cloud, Gaussian Splats, or a combination of both as input, and directly outputs a complete scene graph as a structured JSON script. The graph lists all objects, their semantic attributes, and the relationships between them, while avoiding all of the above mentioned limitations. The choice of multiple input modalities is purely for versatility, allowing a single set of weights to handle diverse scenarios. Point cloud inputs are encoded via Sonata and Gaussian Splat inputs via Chorus, with both modalities projected onto a shared voxel grid and fused through a novel per-voxel contrastive alignment loss before being decoded by a large language model. As a natural consequence of the LLM, GraphWrit3R also supports open-vocabulary querying. On the 3DSSG benchmark, our method achieves state-of-the-art performance on object class, predicate, and triplet recall, outperforming methods that rely on ground-truth object annotations during inference. We further provide qualitative results and analyze different input modality configurations, contrastive loss formulations, and token fusion strategies. Our code and model will be released upon publication.

Existing Driving VLAs predict trajectories while largely ignoring their visual tokens– a phenomenon we trace not to insufficient training but to a structurally ill-posed task formulation. We show that trajectory recovery, when viewed through the lens of inverse kinematics, requires both a current and a future visual state as boundary conditions; existing VLAs supply only the former, which encourages the model to shortcut through ego status and text commands alone. To address this, we re-design Driving VLA in the style of an inverse kinematics solver. First, a next visual state prediction objective that requires the LLM to predict the future visual scene provides dense visual supervision and suppresses shortcut paths. Second, a separate Inverse Kinematics Network (a cross-attention-based conditional diffusion model) that takes only the current and future visual states as input is designed to suppress reliance on ego status and textual shortcuts during trajectory decoding. With this simple prescription alone, our 0.5B-scale model recovers visual grounding and reaches trajectory planning performance comparable to 7B–8B VLAs more than an order of magnitude larger, on both the closed-loop NAVSIM-v2 and the nuScenes benchmarks. Extensive analysis further shows that this improvement stems from a recovered ability to exploit visual features, with the effect being most pronounced in dynamic driving situations such as turning.

Tensors, which give a faithful and effective representation to deliver the intrinsic structure of multi-dimensional data, play a crucial role in an increasing number of signal processing and machine learning problems. However, tensor data are often accompanied by arbitrary signal corruptions, including missing entries and sparse noise. A fundamental challenge is to reliably extract the meaningful information from corrupted tensor data in a statistically and computationally efficient manner. This paper develops a scaled gradient descent (ScaledGD) algorithm to directly estimate the tensor factors with tailored spectral initializations under the tensor-tensor product (t-product) and tensor singular value decomposition (t-SVD) framework. With tailored variants for tensor robust principal component analysis, (robust) tensor completion and tensor regression, we theoretically show that ScaledGD achieves linear convergence at a constant rate that is independent of the condition number of the ground truth low-rank tensor, while maintaining the low per-iteration cost of gradient descent. To the best of our knowledge, ScaledGD is the first algorithm that provably has such properties for low-rank tensor estimation with the t-SVD. Finally, numerical examples are provided to demonstrate the efficacy of ScaledGD in accelerating the convergence rate of ill-conditioned low-rank tensor estimation in a number of applications.

As GUI agents increasingly rely on screenshots to perceive and operate digital environments, they may inadvertently expose sensitive information such as identities, accounts, locations, and behavioral traces. While existing benchmarks primarily focus on task completion, grounding, or defenses against third-party attacks, current visual privacy datasets remain largely restricted to static natural images, limiting their ability to capture the contextual dependence and task relevance of privacy risks in GUI task trajectories. To bridge this gap, we introduce GUIGuard-Bench, a first-step benchmark for studying privacy-preserving GUI agents in trajectory-based GUI workflows. GUIGuard-Bench contains 241 real GUI-agent trajectories with 4,080 screenshots across Android and PC environments. Each screenshot is annotated at the region level with privacy bounding boxes, semantic privacy categories, risk levels, and whether the private information is necessary for completing the task. Built on these annotations, GUIGuard-Bench supports three complementary evaluations: privacy recognition, offline planning fidelity under protected screenshots, and the utility impact of different protection strategies. Our results show that current models can often detect whether a screenshot contains private information, but they struggle with fine-grained localization, category recognition, risk assessment, and task-necessity judgment. We also find that closed-source models, exemplified by Claude Sonnet 4.6, can maintain largely consistent planner semantics in Android environments after privacy protection is applied. GUIGuard-Bench serves as a seed benchmark for studying privacy awareness, data minimization, and privacy-utility trade-offs in GUI agents, and is intended to support broader future evaluation efforts.

Understanding where and why Graphical User Interface (GUI) agents fail is essential for building more reliable systems, yet current evaluation relies on step accuracy, a metric that treats each screen independently and overlooks the underlying structure of GUI environments. This leads to two critical blind spots: (1) functionally equivalent screens are evaluated in isolation, obscuring systematic failure patterns across shared screens; and (2) the long-tailed GUI distribution renders failures on rare but critical screens invisible under standard metrics. To address these issues, we propose \textbf{GUITAR}, a state-centric diagnostic framework that performs structured failure analysis over both states and transitions, using a State Transition Graph (STG) by mapping visually diverse screens to shared functional states. Evaluated with 8 agents across 3 tasks on AndroidControl and 3 tasks on Mind2Web, GUITAR reveals that failures are both structurally concentrated and metrically invisible: 61.9\% of failure instances occur in 20\% states, consistently localizing to a small set of bottleneck states. Furthermore, leveraging bottleneck states for inference guidance yields +2.8\% performance improvement. These findings highlight structure-aware evaluation as necessary for GUI agent diagnosis, and demonstrate that targeted intervention on bottleneck states offers a practical path to improving reliability.

At ultra-low bitrates, high-fidelity reconstruction requires sampling plausible videos from the posterior rather than regressing to oversmoothed conditional means. We propose Generative Video Codebook Codec (GVCC), a zero-shot framework in which a pretrained video generative model serves directly as the decoder, and the transmitted bitstream specifies its generation trajectory. Modern rectified-flow video models are typically sampled with deterministic ODE solvers, which leave no per-step stochastic channel for transmitting compressed information. GVCC addresses this by converting the deterministic flow sampler into an equivalent marginal-preserving stochastic process, so that information can be transmitted by encoding the per-step stochastic innovations. Unlike images, videos introduce longer temporal dependencies and more diverse conditioning modes. We instantiate GVCC in three practical modes: Text-to-Video (T2V) without a reference frame, autoregressive Image-to-Video (I2V) with tail latent correction, and First-Last-Frame-to-Video (FLF2V) with boundary-sharing Group of Pictures (GOP) chaining. On UVG, GVCC achieves the lowest LPIPS among evaluated baselines across three representative bitrate regimes (down to ${\sim}$0.003\,bpp), with 65\% LPIPS reduction over DCVC-RT at matched bitrate.

Hierarchical 3D grouping aims to recover scene groups across multiple granularities, from fine object parts to complete objects, without relying on semantic labels or a fixed vocabulary. The main challenge is to transform 2D foundation-model cues into coherent hierarchy supervision and embed that hierarchy in a 3D representation. We propose H2G, a hyperbolic affinity field for hierarchical 3D grouping. Our method derives semantically organized tree supervision by interpreting foundation-model affinities through Dasgupta's objective for similarity-based hierarchical clustering. This supervision is distilled into a single Lorentz hyperbolic feature field, whose geometry is well suited for tree-like branching structures. A hierarchy-aware objective aligns the field with fine-level assignments, coarse object structure, compact feature clusters, and LCA (Lowest Common Ancestor) ordering. This formulation represents multiple grouping levels in one feature space, enabling semantic hierarchical grouping grounded in 2D foundation-model knowledge.

Hallucination is often viewed as a direct consequence of missing knowledge: a model answers incorrectly when the correct answer is absent from its generation-time distribution, and correctly when it is present. We test this assumption by introducing a semantic notion of answer availability that aggregates token-level variants expressing the same answer concept, and asks whether the correct concept is already available at the moment the model commits to an answer. Across Qwen and Llama models from 0.8B to 72B in both Instruct and Base variants, 16-47\% of Instruct hallucinations occur with substantial probability mass already on the correct concept, and the rate rises monotonically with scale. Comparing such failures against correct generations with matched semantic support, the distinguishing factor is not whether the correct concept is represented, but how its probability is distributed: correct generations concentrate mass on a single surface form, hallucinations disperse it across alternatives. The same sharpening asymmetry extends across multi-token generation and is detectable in pre-generation hidden states. Together, these results identify a single mechanism: instruction tuning sharpens answer commitment with scale, making helpfulness and confident hallucination two consequences of the same underlying disposition.

Visual geometry grounded transformers are strong feed-forward backbones for multi-view 3D reconstruction, but their global attention becomes the dominant inference bottleneck as views and tokens increase. Existing acceleration methods mainly focus on designing compression operators, while compression strengths are often assigned by uniform schedules or offline calibration, overlooking the heterogeneous sensitivity of layers and attention heads during inference. We therefore propose HALO-VGGT, a lightweight online compression allocator requiring no training or calibration for efficient VGGT-style reconstruction. HALO uses a sampled query-key margin proxy to estimate compression sensitivity at the layer and head levels, and assigns hierarchical compression ratios accordingly. Instead of introducing a new compression operator, HALO complements pruning, merging, and sparse attention mechanisms by deciding where and how aggressively they should be applied. HALO improves the trade-off between speedup and accuracy across datasets and backbones, achieving up to 4.20$\times$ speedup and further improving existing merging and sparsity methods by up to 1.46$\times$.

Periodic structure is a fundamental basis for long-term time series forecasting (LTSF), but real-world periodic patterns are not simple repetitions. Variations in level, amplitude, and phase within one period can accumulate across subsequent periods and reshape future periodic structures. We therefore argue that explicitly modeling inter-period relations is a core challenge in LTSF. Existing methods often handle these relations implicitly or rely on unconstrained mixing, where meaningful dependencies can be entangled with noisy interactions. Simply increasing mixing expressiveness does not resolve this issue. Motivated by this, we address LTSF through gradual state refinement driven by interactions among periods. The key challenge is to reinforce meaningful inter-period dependencies while suppressing noise and maintaining stable representations during repeated refinement. To this end, we draw inspiration from Hamiltonian dynamics and use its split-state formulation and symplectic-style updates as an architectural bias for stable iterative refinement. We propose HAMSTAR (HAMiltonian STructured period-Aligned Representation for time series), a Hamiltonian-inspired framework for structured period-aligned time series forecasting. HAMSTAR represents time series as period-aligned latent states, decomposes them into split-state components, and uses structured inter-period coupling to transform period relations into refinement signals. Symplectic-style updates then progressively refine these states, enabling stable representation dynamics. Experiments on standard LTSF benchmarks show that HAMSTAR achieves state-of-the-art performance across diverse datasets, and further analyses demonstrate meaningful inter-period modeling and stable latent dynamics.

Multi-agent LLM and VLM systems introduce a fragmentation blind spot: an adversary can decompose a harmful objective into individually benign sub-tasks distributed across agents, tools, and shared memory, so that each local interaction appears policy-compliant while the composed workflow is unsafe. Existing defenses classify isolated prompts, monitor single-agent histories, or impose architectural constraints, but none directly detect runtime system-level drift toward harm across the full interaction graph. We introduce CITADEL (Composite Intent Tracking for Agentic Defense via Evolving Latent States), a training-free runtime monitor for multi-turn, multi-agent jailbreaks. CITADEL maintains latent states over agents, tools, and memory stores using a cosine-gated recurrence with no learned parameters; encodes a published safety taxonomy as fixed harm anchors via the same frozen multimodal encoder applied to runtime events; and scores risk through the conjunction of harm-anchor proximity, multi-node participation, and positive temporal drift. We evaluate CITADEL on MA-SafeBench, a new benchmark spanning five multi-agent attack families, three communication topologies, and heterogeneous API-accessible LLM/VLM backbones. CITADEL reduces average attack success rate from 69.3% to 16.9% — a 75.6% relative reduction — at a 2.1% false-positive rate and 28 ms per-event overhead, outperforming per-agent detectors, trajectory-level monitoring, and architectural defenses. The same hyperparameters transfer without retuning to OpenAgentSafety, Agent Security Bench, and MTMCS-Bench, indicating that collective geometric convergence provides an effective runtime signal for detecting distributed jailbreaks in multi-agent systems.

Audio Language Models (ALMs) have recently shown strong capabilities in unified reasoning over speech, sound, and natural language; yet we find that they can inherit sycophancy, the tendency to agree with user assertions even when they contradict objective evidence. This failure mode is especially concerning for audio-conditioned reasoning, where a model must preserve evidence from acoustic events, speaker characteristics, and speech rate while responding to potentially misleading user feedback. However, unlike text and vision-language sycophancy, ALM sycophancy has not been systematically studied. We therefore introduce SYAUDIO, the first benchmark dedicated to evaluating sycophancy in ALMs, consisting of 4,319 audio questions spanning Audio Perception, Audio Reasoning, Audio Math, and Audio Ethics. Built upon established audio benchmarks and augmented with TTS-generated arithmetic and moral reasoning tasks, SYAUDIO enables systematic evaluation across multiple domains and sycophancy types with carefully verified data quality, including a human-speaker validation of the TTS pipeline. Using this benchmark, we identify substantial and audio-specific sycophancy patterns under realistic conditions involving noise and speech rate, and further show that supervised fine-tuning reduces misleading susceptibility while decode-time steering reveals controllable hidden-state directions for both misleading susceptibility and correction receptiveness.

Proactive task-oriented dialogue (TOD), such as outbound sales, demands a persuasive agent that actively probes the user's concerns and steers the conversation toward acceptance within a bounded number of turns. Yet post-trained LLMs are inherently conservative, and reward-shaping RL (e.g., GRPO) struggles since it only re-weights what an already passive policy samples. We show that conditioning on the user's latent concerns unlocks proactive capability that no amount of sampling can undermine, establishing these concerns as a pivotal training-time signal. To operationalize this finding, we build the \textbf{Cognitive User Simulator}, which models each user as a stratified persona comprising observable external traits and hidden internal concerns. The simulator produces faithful and diverse interactions, while emitting per-turn state dynamics that track persuasion progress. We then introduce \textbf{Simulator-Induced Asymmetric-View Policy Optimization}, which converts the modeled concerns and the simulation state transition into complementary training objectives: (1) \emph{Asymmetric On-Policy Self-Distillation} that transfers concern-aware behavior from a privileged view of the same policy into its deployable, conversation-only view; and (2) \emph{State-Transition Policy Refinement} where the final decision provides trajectory-level advantage while synchronous state transitions refine turn-level credit direction. Across two real-world food-delivery benchmarks (\emph{merchant} and \emph{courier} outbound recruitment), our method matches or surpasses leading proprietary LLMs, outperforms strong RL baselines, and generalizes across various LLM-based user simulators.

Embodied agents make decisions by combining information from heterogeneous memory modules, including declarative memory for explicit knowledge, procedural memory for learned routines, and working memory for current observations. During interaction, the memory contents required for decision making change with the agent's state, goal, and task progress, while many contents remain shared across consecutive decisions. However, existing memory modules are typically accessed through separate interfaces, causing each updated memory context to be processed as a new input to the foundation model. This prevents selective key-value (KV) cache reuse for unchanged memory contents and forces full KV cache recomputation even when much of the memory context remains valid. We introduce HeMeR, an inference framework that performs KV reconciliation between dynamically updated memory contexts and reusable KV states. HeMeR tracks declarative, procedural, and working memory contents across action decisions, preserves KV states for unchanged contents, and reconciles only the states affected by memory-content updates. This enables efficient and stable inference as memory contents evolve during interaction. We evaluate HeMeR on ALFWorld, Habitat, and real-world robotic deployments requiring coordinated use of heterogeneous memory contents. On Real-world, HeMeR achieves $83.20\%$ task success and reduces per-step inference time from $2.94$ to $1.87$ seconds compared with MemRL by preserving compatible KV states as task progress updates the memory context.

Federated continual learning (FCL) evaluates how distributed clients learn from changing data streams while retaining previously learned knowledge. Existing evaluations are difficult to compare because they often change datasets, task splits, client data splits, task orders, backbones, memory assumptions, and reporting rules simultaneously. We introduce \textbf{HERO}, a heterogeneity-aware benchmark library for FCL. HERO builds benchmark streams by separating three choices that are often coupled, namely the task split, the client data split, and the client task sequence. In HERO-Core, the main comparable benchmark, $\alpha$ controls client data skew and $\rho$ controls task-order mismatch. We evaluate representative FCL methods on CIFAR-100 and TinyImageNet using final average accuracy, average forgetting, and bottom-10\% client accuracy. We also include a graph-based Domain-IL portability case study on OGB-MolPCBA, where scaffold-domain granularity changes the input distribution while the prediction task remains fixed. Our results show that method behavior changes across easy and heterogeneous settings, that average accuracy can hide weak bottom-client performance, that task-order mismatch favors different strategies from synchronized evaluation, and that the same HERO interface can expose domain-shift difficulty beyond image-based FCIL. HERO releases benchmark streams, configurations, method implementations, and reporting scripts to support reproducible and setting-aware FCL evaluation.

We present a simple, inexpensive, and effective method for heteroscedastic uncertainty quantification in neural networks. We build on Variational Bayesian Last Layers (VBLL), wherein deterministic training objectives are developed for variational inference of the network last layer. In particular, we (1) Introduce t-VBLL layers, which perform variational inference for the aleatoric noise covariance, and (2) Introduce Het-VBLL, a Bayesian last layer scheme to model heteroscedastic noise. These methods are based on novel, analytically tractable evidence lower bounds. We further discuss parameterization and initialization within these models. We show that these novel design elements enable effective uncertainty modeling at minimal additional cost, and substantially improve performance over similar methods such as VBLLs.

Forecasting the behavior of spatio-temporal systems often requires more than accurate predictions: models should uncover the dynamical modes governing future evolution. Existing temporal graph and latent dynamical models achieve strong forecasting performance, but their latent representations often entangle multiple dynamical patterns, limiting interpretability. Conversely, regime-inference methods explicitly model such modes but typically do not support generative forecasting over graph-structured systems. We introduce HERON, a self-supervised regime-aware generative framework that unifies graph-structured trajectory generation with regime identification. HERON factorizes the latent state into discrete system-level regimes, capturing global dynamical modes, and continuous entity-level dynamics, yielding a modular regime interface compatible with existing spatio-temporal backbones. Crucially, the inferred regime actively conditions system dynamics, enabling regime tracking and controlled simulations under alternative dynamical modes. Without regime annotations, HERON jointly learns forecasts and regime dynamics in a self-supervised manner, recovering meaningful modes across controlled synthetic and real-world datasets while maintaining competitive forecasting performance. These results open a path toward interpretable, controllable generative modeling of spatio-temporal dynamics.

Open-vocabulary indoor scene synthesis aims to generate plausible layouts from arbitrary user instructions while ensuring physical feasibility and semantic consistency. Existing methods directly infer spatial relations between objects from instructions and solve for layouts based on predefined rules, achieving progress in physical feasibility. However, they often struggle to precisely align with complex instructions due to insufficient comprehension of spatial relations and limited capability to translate them into layouts. In this paper, we propose a Hierarchical Reasoning and Reflection agent framework (HierRR), which leverages a relational hierarchy to enable multi-granularity spatial relation comprehension and attributable layout reasoning, improving semantic consistency while maintaining physical feasibility. Specifically, HierRR introduces the spatial transformation chain‑of‑thought reasoning module that explicitly models the mapping from spatial relations to executable geometric placements, mitigating semantic drift during reasoning. Furthermore, the semantic- and vision-guided reflection module is employed to iteratively refine spatial relations and layout reasoning from local to global along the hierarchy, effectively bridging the gap between language instructions and geometric layouts. Extensive comparative experiments and ablation studies demonstrate that HierRR generates more plausible and semantically consistent indoor scene layouts than existing methods, while generalizing effectively to complex scenes with numerous objects and diverse spatial constraints.

Training large foundation models at low numerical precision is one of the most promising directions for reducing the compute and memory cost of modern AI. Recent 4-bit floating-point formats such as MXFP4 and NVFP4 can be applied to linear GEMM operations in LLMs, but their limited dynamic range introduces numerical instability that prior work addresses by stacking stabilization mechanisms, typically executed at higher precision and partially eroding the efficiency gains that motivate FP4. In this work, we argue that numerical format design is itself a first-class lever for stable FP4 training, and present the first systematic study of FP4 LLM pretraining on energy-efficient Huawei Ascend NPUs. We compare the recently proposed HiFloat4 (HiF4) format against MXFP4 across both dense (OpenPangu-1B, Llama3-8B) and Mixture-of-Experts (Qwen3-MoE-30B) architectures, executing all linear and expert GEMMs in FP4. HiF4's hierarchical scaling provides enough representational headroom that a single stabilization step suffices to keep the relative loss within 1\% of a full-precision baseline — roughly half the gap of MXFP4, which requires three such mechanisms — while incurring less than 1\% average degradation on downstream tasks. Our results suggest that stable, accurate FP4 training does not require an ever-growing stack of stabilization techniques; it requires the right numerical format.

While attention mechanisms are the cornerstones of modern foundation models, the theoretical understanding remains largely elusive. In this paper, we analyze the attention-indexed model, a rich framework that encapsulates multi-layer and multi-head attention. We first prove that while, in a suitable high-dimensional limit, the macroscopic landscape of the population loss is characterized by a finite set of order parameters, the corresponding population gradient flow unfolds in an infinite-dimensional state space, which we show can be exponentially well-approximated by a finite truncated system. Our dynamical analysis uncovers how standard attention parameterizations act as architectural implicit biases to overcome the curse of the information exponent—a barrier that typically traps the direct optimization of the attention matrix $S$. Specifically, the tied attention ($S=WW^T$) induces an automatic symmetry-breaking geometry at initialization, shielding the gradient flow from uninformative manifolds and achieving weak recovery in $\Theta(1)$ time. For untied attention ($S=UV^T$), we reveal a timescale separation dynamic: efficient weak recovery hinges on whether the fast-timescale evolution of the pre-activation mean successfully breaks initial symmetries.

The trend towards larger training setups has brought a renewed interest in partially asynchronous two-phase optimizers which optimize locally and then synchronize across workers. Additionally, recent work suggests that the one-worker version of one of these algorithms, DiLoCo, shows promising results as a (synchronous) optimizer. Motivated by these studies we present an analysis of LA-DiLoCo, a simple member of the DiLoCo family, on a high-dimensional linear regression problem. We show that the one-worker variant, LA, provides a different tradeoff between signal and noise than SGD, which is beneficial in many scenarios. We also show that the multi-worker version generates more noise than the single worker version, but that this additional noise generation can be ameliorated by appropriate choice of hyperparameters. We conclude with an analysis of SLA --- LA with momentum --- and show that stacking two momentum operators gives an opportunity for acceleration via a non-linear transformation of the ``effective'' Hessian spectrum, which is maximized for Nesterov momentum. Altogether our results show that two-phase optimizers represent a fruitful new paradigm for understanding and improving training algorithms.

Reconstructing lineages from live-imaging microscopy requires linking cell detections across time, including through cell divisions. A common approach is to construct a candidate graph and associate cell segmentations (nodes) across frames. However, these and other existing methods overlook two structural obstacles in candidate tracking graphs: (i) cell divisions entangle distinct lineage paths in the node embedding space, and (ii) edges sharing a node have near-random label agreement, so the candidate-graph topology carries no useful information for graph neural networks to aggregate. We propose the Higher-Order Cell Tracking Transformer (HOCT), an edge-centric architecture in which candidate cell links attend to one another under a 3D geometric prior, resolving both issues. Evaluated on the Cell Tracking Challenge and a bacteria division benchmark, HOCT achieves state-of-the-art results without deep pre-trained image encoders. Moreover, the proposed approach is easier to fine-tune, quickly reducing tracking errors by 59% with 400 annotations in a human-in-the-loop setting, outperforming LoRA fine-tuning of competing transformer baselines (6.75% improvement).

The advancement of graph-based malware analysis is critically limited by the absence of large-scale datasets that capture the inherent hierarchical structure of software. Existing methods often oversimplify programs into single-level graphs, failing to model the crucial semantic relationship between high-level functional interactions and low-level instruction logic. To bridge this gap, we introduce HiGraph, the largest public hierarchical graph dataset for malware analysis, comprising over 200M Control Flow Graphs (CFGs) nested within 499K Function Call Graphs (FCGs). This two-level representation preserves structural semantics essential for building robust detectors resilient to code obfuscation and malware evolution. We demonstrate HiGraph’s utility through a large-scale analysis that reveals distinct structural properties of benign and malicious software, establishing it as a foundational benchmark for the community. The dataset and tools are publicly available at https://higraph.org.

Millimeter-wave radar point clouds provide a visually privacy-preserving representation for human motion analysis, but object-related interactions remain difficult to describe because the object evidence that defines the interaction is not always separable or consistently observable in sparse radar returns. We address this problem without direct object detection. Instead, we infer hidden interaction semantics from the temporal structure of observation-conditioned corrections to predicted body motion, motivated by the view that object affordances shape how the body moves. To this end, we propose HIRaD, a radar semantic sensing framework that organizes sparse radar returns into a predictive node-structured latent state of local motion carriers. It represents interaction-relevant cues as corrective residual trajectories between a motion-continuation prior and an observation-conditioned posterior, and compresses these trajectories into a language-aligned semantic bottleneck. A structured prefix conditions a frozen autoregressive language model on both the semantic bottleneck and node-level motion summaries, separating compact interaction-level cues from node-level motion evidence. Experiments show that our proposed method improves radar captioning and interaction understanding under weak object observability over strong baselines, while ablations validate the importance of radar-conditioned prefixing, latent state representation, and residual trajectory modeling.

The deployment of large language models (LLMs) in hiring pipelines is growing, yet existing evaluations generally overlook an important aspect of this development: the compliance of LLMs with applicable legal rules. To fill this gap, we introduce HireWatch, a benchmark that assesses whether LLM hiring assistants comply with U.S. employment discrimination law (e.g., the Civil Rights Act of 1964). We construct a 200-item bank consisting of interview questions that are either lawful or unlawful under U.S. federal statute, regulation, and enforcement guidance. We evaluate the propensity of 18 proprietary and open-source LLMs to select lawful vs. unlawful questions from this bank across ~100,000 trials under a novel seven-tier contextual escalation ladder, crossing occupational details with organizational pressure. Our findings reveal that LLMs' compliance with employment discrimination law is brittle. While LLMs ordinarily comply with legal rules under different job contexts, they readily violate the law when subject to institutional pressure (e.g., an unlawful manager instruction or company policy). Meanwhile, although the inclusion of explicit legal reminders in prompts can improve legal compliance, it does not altogether eliminate violations of employment discrimination law.

Scaling laws govern macroscopic resource allocation for LLMs, yet precise architectural configurations for Mixture-of-Experts (MoE) models remain guided by heuristics. Existing MoE scaling studies either incorporate MoE-specific variables into scaling formulas, causing fitted coefficients to grow rapidly without proportional experimental support, or fix all non-MoE factors, implicitly assuming global architecture does not influence local MoE scaling. We propose a reusable framework for holistic MoE architectural optimization. We first reveal that relying solely on FLOPs per token ($M$) biases MoE evaluation, as heterogeneous Attention/FFN computational densities enable parameter inflation without effective compute gains. We therefore establish a joint constraint triad of $M$, active parameters ($N_a$), and total parameters ($N$) for rigorous MoE characterization. To tame the resulting $\mathcal{O}(n^{16})$ search space, we employ mathematical decoupling: structural constraints and a rank-preserving property of the hidden dimension factorize the optimization into an $\mathcal{O}(n^3) + \mathcal{O}(n^2)$ two-phase search. Through extensive validation across 670+ MoE models spanning $10^{18}$ to $3 \times 10^{20}$ FLOPs, we derive globally applicable scaling laws mapping any compute budget to its optimal architecture. A key finding is that the near-optimal configuration band widens with scale, providing a quantitative basis for trade-offs between scaling law recommendations and engineering constraints. Our work delivers actionable blueprints for optimal MoE design under arbitrary compute budgets.

Reconstructing a dynamic 4D scene along a novel camera trajectory requires both metric geometry from the source video and generative completion of disoccluded target-view regions. Feed-forward 4D reconstruction models recover increasingly comprehensive input-view geometry, including depth, point maps, and dense 3D tracks, but remain tied to the observed camera stream. Camera-controlled video diffusion models (VDMs) provide strong generative priors for novel-view synthesis, yet their camera control is typically not precise enough for metric novel-view reconstruction. We argue that comprehensive input-view 4D reconstruction is an effective geometric interface between these two capabilities: properly exposed to a VDM, it turns reconstruction outputs into camera-accurate novel-view RGB videos that remain geometrically consistent under downstream 4D reconstruction. We introduce HOLO4D, a geometry-aware video-to-video framework that controls a pretrained VDM with a hybrid 4D geometric cache built from the source video. A small set of source views covering the camera motion forms the static cache, while dynamic regions are rendered from the matching source frame. The resulting hybrid RGB-D representation is combined with explicit valid masks, target-camera ray embeddings, and dense 3D tracking tokens. Conditioned on this 4D scaffold, the diffusion backbone synthesizes the target-view RGB video. Under a common downstream 4D reconstructor, our generated videos yield target-view depth more consistent with the held-out reconstruction than that of the baselines, and an additional feature-level readout shows that intermediate VAE features already encode target-view depth. Experiments and ablations further show that progressively richer 3D/4D reconstruction signals improve camera adherence.

Large language mode agents are moving beyond text-only interaction toward physical-world control, with smart homes as a representative domain. Real domestic interaction requires understanding ambiguous intents, operating in dynamic environments, and performing multi-turn reasoning. However, existing methods struggle to generate high-quality training data for smart home agents. We propose HomeFlow, a verifiable data flywheel for this domain. HomeFlow uses HomeEnv as a unified simulation environment and HomeMaker to procedurally generate diverse home settings. Subsequently, Blueprint compiles open-ended user intents into executable state-based success conditions, while MCTS-Flow synthesizes diverse, verifiable multi-turn trajectories through environment-guided tree search. We then optimize the agents via supervised fine-tuning and step-wise RLVE, which facilitates iterative improvement through authentic physical feedback. We further construct SmartHome-Bench to evaluate the agent across various smart home tasks. On this benchmark, HomeFlow-RL-4B and HomeFlow-RL-8B achieve task success rates of 84.60\% and 87.03\%. It is worth noting that HomeFlow-RL-8B even surpasses the leading GPT-5.5 by 1.23 percentage points.

Learning accurate value functions plays a decisive role for reinforcement learning (RL) agents to solve long-horizon, complex tasks. Conventional temporal-difference (TD) learning objectives suffer from value-estimation bias that accumulates over the horizon, while extended-horizon modeling methods, such as $n$-step TD backups and Q-chunking, adopt a rigid, fixed-horizon value-modeling recipe that is often not flexible enough to capture complex value structures in long-horizon, multi-stage tasks. In this paper, we show that enabling value updates with dynamic horizon composition can yield a strong offline policy learning scheme. Our method, _Horizon Adaptive Offline Policy Learning via **VA**lue **ST**itching_ (**_VAST_**), replaces fixed-horizon backups with recursive, horizon-adaptive value composition. Its key ingredient is to couple value optimization with a future state- and horizon-length-conditioned **_auxiliary value function_** that is learned through direct data supervision, and a **_stitching policy_** that optimally selects the reward-maximizing horizon length and future sub-goal to achieve horizon-adaptive value stitching. This design enables direct estimation and compositional “stitching” of variable-length returns grounded in actionable sub-goal states, providing an accurate and greedily exploitable value-supervision signal for offline policy optimization. Across 50 tasks on OGBench, _VAST_ outperforms fixed-step, extended-horizon methods, and generative-value offline RL baselines, achieving strong performance particularly in high-complexity, long-horizon decision-making tasks.

In artificial and biological neural networks, concepts are often encoded as consistent linear directions in representation space. In deep learning, this idea is known as the linear representation hypothesis and underpins many interpretability and control methods based on linear probes, from concept detection to activation steering. Yet while prior work has studied whether such directions should exist \textit{after} training, the dynamics of how they emerge \textit{during} training remain poorly understood. Here, we develop a framework to study the alignment of concept directions during training -- a process we call "abstraction". In a minimal linear network setting, we obtain exact solutions for the full trajectory of abstraction. These solutions reveal key analytic principles governing abstraction: (i) data and target geometry jointly determine terminal abstraction, (ii) abstraction improves with network depth, and (iii) initialization scale controls the maximum abstraction reached during training. Extending our theory to nonlinear networks, we analyze how the choice of nonlinearity affects abstraction dynamics: erf networks approximate the linear theory, while abstraction in ReLU networks depends less on target geometry and more on input geometry. Across both, we prove a striking attenuation law: both nonlinearities weaken abstraction in activations relative to preactivations. We find evidence for this law in open models (DINOv3, Gemma 4) and apply our theory to improve linear probe generalization in LLMs. Together, our results provide a dynamical theory of abstraction with implications for interpretability and control.

Tool calling, the ability to invoke external tools on demand, is central to agentic LLMs. Within an agent, the LLM decides whether and when to use these tools, greatly extending the agent’s capability boundaries. However, the mechanism that determines whether a model calls a tool or responds directly remains poorly understood. Agentic prompts are often long and heavily scaffolded, combining role instructions, tool schemas, format templates, and user requests across thousands of tokens. This creates a noisy and highly entangled context that makes it difficult to identify a single controllable variable for mechanistic analysis.To obtain such a variable, we propose a method to convert complex agentic prompts into minimal contrastive pairs, in which a single request verb determines the tool-call decision. Replacing an execution verb, such as \textit{write}, with an analysis verb, such as \textit{discuss}, reliably flips whether the model chooses to call a tool. This suggests that the choice is mediated by a compact internal state. 1,500 paired prompts across Python, Java, and C++, where 1,200 are for mechanistic analysis, and 300 are held out for evaluation, are constructed.We trace the tool-calling decision to a single vector, $\mu_\Delta$, and show that it is both causally necessary and sufficient.Transcoder analysis reveals how this vector forms: the scaffold makes tool calling the default, while analysis verbs suppress this default by activating features that signal no tool is needed. Execution verbs activate no analogous tool-use features, so $\mu_\Delta$ reflects the scaffold-induced default rather than a positive execution-verb signal. Downstream attention and MLP layers then read out this vector into the first output token. The mechanism replicates across diverse model families. Our code is available at \url{https://anonymous.4open.science/r/MI4ToolCalling}.

While large language models (LLMs) appear to be increasingly capable of solving compositional tasks, it is an open question whether they do so using compositional mechanisms. In this work, we investigate how feedforward LLMs solve two-hop factual recall tasks, which can be expressed compositionally as $g(f(x))$. We first confirm that modern LLMs continue to suffer from the "compositionality gap", i.e. their ability to compute both $z = f(x)$ and $y = g(z)$ does not entail their ability to compute the composition $y = g(f(x))$. We then decode residual stream representations and identify two processing mechanisms: one which solves tasks *compositionally*, computing $f(x)$ along the way to $g(f(x))$, and one which solves them *directly*, without any detectable signature of the intermediate variable $f(x)$. Finally, we find that embedding space geometry is strongly related to which mechanism is employed, where the idiomatic mechanism is dominant when tasks are represented by translations from $x$ to $g(f(x))$ in the embedding spaces.

Vision models are sensitive to object-background context, yet it remains unclear how recognition changes as the semantic distance between objects and backgrounds increases. We introduce ImageNet-OOC1k, a benchmark enabling continuous control of object-background relationships via a consensus semantic dissimilarity axis. This formulation reveals a highly consistent monotonic decline in recognition accuracy across 35 models spanning convolutional neural networks, Vision Transformers, and vision-language models. Internal analyses of Vision Transformers reveal aggregation failure, where object evidence is present at the patch level but not consolidated into the final prediction. Leveraging this characterization, a simple spatial re-ranking strategy recovers a subset of these errors and improves accuracy by up to 5.4 percentage points. These results establish context sensitivity as a continuous and predictable property of modern vision systems, with Vision Transformer analyses showing that failures can arise from limitations in aggregating competing signals rather than from missing object representations.

In social choice, anonymity (treating all agents equally) and neutrality (treating all alternatives equally) are widely regarded as "minimal demands" and "uncontroversial" axioms of equity and fairness. However, due to the ANR impossibility theorem, no voting rule can simultaneously satisfy anonymity, neutrality, and resoluteness (always choosing a unique winner). While the worst-case ANR impossibility is well understood, its likelihood in probabilistic models remains underexplored. We address this question through comprehensive likelihood analysis under semi-random models, providing accurate bounds to quantify the advantage of the optimial tie-breaking mechanisms over other commonly-used tie-breaking mechanisms. Our characterizations reveal an $n^{\Theta(m!)}$ improvement in many cases, where $n$ is the number of agents and $m$ is the number of alternatives. We also prove a quantitative ANR impossibility theorem characterizing the tradeoff between equity (anonymity and neutrality) and a new efficiency criterion called $\delta$-anonymity. Our work advances the understanding of trade-offs between equity and efficiency in voting.

Evaluating general-purpose multimodal large language models (MLLMs) for surgical scene understanding remains limited by scarce surgical data and fragmented evaluation protocols. We address these challenges through a two-stage evaluation framework for surgical scene understanding. In Stage I, we introduce SURGKNOWBENCH, a diagnostic benchmark that harmonizes public surgical datasets across procedures and tasks and evaluates proprietary and open-weight MLLMs to map the boundary of pretrained surgical competence. SURGKNOWBENCH reveals that current models retain partial visual priors but remain weak on procedural reasoning and workflow assessment. Building on this failure frontier, and motivated by the scaffolded nature of surgical training, we introduce SURGAGENTGYM, a controlled environment where general-purpose MLLM agents can use bounded knowledge search over SURGWIKI and general-purpose visual grounding for surgical tasks. SURGAGENTGYM evaluates whether generalist agents can improve surgical understanding at inference time through tool-mediated evidence use, without additional surgical training. Our results show that agentic inference narrows the gap mainly for perception-heavy tasks, while fine-grained localization and expert-style assessment remain persistent bottlenecks. This opens a path toward future studies of agentic surgical understanding. We have released the codebase and artifacts.

Self-supervised multimodal representation learning has achieved remarkable success across diverse domains, yet capturing synergistic information remains challenging due to the complexity of cross-modal interactions. Unlike the shared information across individual modalities, synergy arises when task-relevant signals emerge only from the joint configuration of multiple modalities and cannot be recovered from any modality in isolation. This work focuses on how to preserve the information capacity for such synergistic signals in multimodal representations. The key observation is that synergistic information is reflected in higher-order statistical dependence among modalities, which provides a principled target for explicitly modeling joint interactions. Motivated by this insight, we propose Higher-order Representation and Information Learning (HRIL), which constructs an empirical cross-moment tensor over modality embeddings to represent multi-way interactions. HRIL employs Tucker decomposition to disentangle non-separable interactions into a core tensor, complemented by a synergy-aware regularizer that prevents energy concentration and preserves higher-order coupling for synergistic information capture. Experiments on synthetic and real-world benchmarks demonstrate consistent improvements over existing multimodal contrastive methods, with notable gains on tasks dominated by synergistic interactions.

Large language models (LLMs) are increasingly adopted for software and hardware design, yet these domains are still evaluated separately. Software benchmarks typically assume fixed hardware targets, while hardware benchmarks focus on component-level optimization without considering the full hardware-software stack. Consequently, no existing benchmark evaluates whether an LLM agent can perform end-to-end, system-level hardware-software co-design. Such a process requires: 1) analyzing applications to identify kernels requiring acceleration, 2) designing and \emph{integrating} heterogeneous accelerators into a System-on-Chip (SoC) under resource constraints, and 3) mapping kernels onto the generated accelerators. We present \hsco{}, an end-to-end hardware-software co-design benchmark for accelerator-rich heterogeneous SoC generation. Built upon an open-source SoC platform with a curated repository structure, HSCO-Bench evaluates the ability of LLMs to jointly optimize software and hardware stacks, producing SoC prototypes deployed on the AMD Virtex-7 FPGA VC707 Evaluation Kit. Experimental results show that end-to-end integration remains challenging for current models, with widespread failures. Among the five frontier models evaluated, only two of them could successfully generate valid SoC prototypes. Yet, even in these successful instances, the generated designs are far from optimal. While we observe a promising peak speedup of 16.22x, the maximum additional resource utilization reaches only 23.67%. This highlights that while state-of-the-art models demonstrate an emerging capability for hardware acceleration, they still heavily underutilize the available hardware capacity, leaving substantial room for future optimization. To the best of our knowledge, HSCO-Bench is the first benchmark targeting this complete co-design flow, enabling LLMs to jointly reason about and modify both the software and hardware stacks of heterogeneous SoCs.

Humanoid robots promise whole-body interaction in human-centered environments, but scalable policy learning remains difficult because task-level semantic reasoning and whole-body level dynamic execution are tightly coupled. A practical solution is hierarchical control, where a high-level policy predicts intermediate whole-body actions and low-level general motion trackers (GMTs) execute them as stable humanoid motion. However, existing benchmarks rarely evaluate the policy–tracker interface itself, leaving open whether intermediate whole-body actions are executable, robust under task distribution shifts, and transferable across different GMT backends. We introduce \textsc{HumanoidArena}, a simulation-first benchmark for egocentric hierarchical whole-body learning. The benchmark formulates policy learning as a hierarchical decision making problem: a high-level policy converts egocentric vision, proprioception, and instructions into a compact whole-body action, which is subsequently executed by a low-level GMT. Instead of treating the legs as planar transport tools, \textsc{HumanoidArena} emphasizes interactions where lower-body coordination is structurally necessary in task completion. We therefore design 7 leg-critical HOI/HSI tasks in which success requires foot placement, balance maintenance, posture adjustment, and whole-body reorientation. To further diagnose the hierarchical system, we evaluate policies from two complementary perspectives: perturbation-conditioned generalization and GMT-conditioned transfer. We benchmark representative imitation-learning and VLA-style policies under this shared interface. Experiments show that hierarchical control enables learned policies to solve diverse leg-critical interactions, but performance is strongly tracker-conditioned and cross-GMT transfer remains fragile. These results position \textsc{HumanoidArena} as a benchmark for studying transferable intermediate action representations and scalable egocentric whole-body policy learning. Code and data are available at: \url{https://anonymous.4open.science/r/HumanoidArena-D36F}.

Recent advances in model architectures, compute, and data scale have driven rapid progress in video generation, producing increasingly realistic content. Yet, there is no method to systematically assess the fidelity of human figures generated in these videos. In this paper, we present HumanScore, a systematic framework to evaluate the \textbf{quality of human figures and motions} in AI-generated videos. We introduce seven metrics spanning kinematic plausibility and biomechanical consistency. Through carefully designed prompts, we elicit a broad set of movements at varying intensities. We evaluated a total of 840 videos generated by 6 state-of-the-art models. Our framework reveals consistent gaps between visual plausibility and motion fidelity, highlights common failure modes, and ranks models from multiple quantitative and physically meaningful metrics. The proposed metrics are correlated with human evaluations and are simple to compute using standard pose estimators.

Infrared-Visible image fusion (IVIF) is essential for robust perception, yet existing methods remain limited under complex degradations. While diffusion models (DMs) offer powerful modeling capabilities to combat such degradations, their iterative process incurs high computational cost. One-step distillation effectively accelerates inference but typically leads to over-smoothed results. To resolve these limitations, we propose a one-step fusion framework with hybrid reinforcement learning, termed OneHRL. We treat degraded inputs as intermediate noisy states along a generative trajectory, enabling a distilled one-step generator to rectify them directly to the clean fusion results. Concurrently, to mitigate the over-smoothed results in distillation, we incorporate the Flow-GRPO paradigm to introduce stochasticity into the deterministic one-step generation. This restores the stochastic exploration necessary for sampling-based reinforcement learning. Furthermore, we design a coarse-to-fine reward that integrates a global semantic reward with an object-level reward. This hybrid reward ensures that the generator captures both naturalness and fine-grained texture. Extensive experiments demonstrate that our method achieves superior fusion quality and efficiency.

Language-conditioned manipulation requires skill representations that are both multi-modal and predictable: the same instruction may admit diverse executions, while the high-level planner must select a stable skill for low-level control. However, vector-quantized skill codes compress continuous variations into isolated symbols, introduce biased straight-through gradients, and lack geometry for relating semantically similar skills. We propose $\textbf{HyperSkill}$, a hierarchical skill-learning framework that models skills as continuous embeddings on the unit hypersphere. We use a learnable von Mises--Fisher mixture prior to capture skill-level multi-modality, with geodesic repulsion encouraging diverse mode coverage on the hyperspherical manifold. To connect training-time skill inference with inference-time planning, we introduce Geodesic Skill Alignment (GSA), which aligns posterior skills inferred from demonstration segments with prior skills predicted by a high-level planner using squared geodesic distance. Together, these components provide a geometrically structured skill interface that preserves execution diversity while enabling causal skill prediction at inference time. For low-level control, a Conditional Flow Matching generator produces smooth, multi-modal action trajectories conditioned on skill embeddings. Experiments on LOReL and Kitchen demonstrate that HyperSkill achieves state-of-the-art success rates on the evaluated benchmarks. Ablation studies further validate the importance of each proposed component.

Temporal action detection in long untrimmed videos still suffers from inaccurate temporal localization and low-quality proposals, especially under sparse temporal annotations and ambiguous event boundaries. In the context of open-ended video understanding and LLM-driven event analysis, reliable temporal action proposals have emerged as a critical interface bridging temporal localization, event annotation, and semantic understanding. However, conventional temporal action detection (TAD) pipelines lack explicit modeling of proposal localization quality, cross-representation support, and inter-proposal relations, making it difficult to consistently provide high-quality proposals. We propose Hybrid Rank-Calibration of Action Proposals (HyrCap), a proposal-centric query-based TAD framework for temporal event understanding. In HyrCap, Hybrid denotes the use of complementary video features that provide different temporal localization cues, while Rank-Calibration refers to calibrating the ranking confidence of dense action proposals so that it better reflects proposal localization reliability and candidate competition. HyrCap treats dense query-based predictions as an over-complete temporal event hypothesis space and improves the ranking reliability, localization quality, and candidate discrimination of action proposals through modules such as Cross-Representation Consensus Calibration, Localization Quality Estimation, and Proposal Duplication Suppression. Furthermore, inspired by the strong capability of LLMs in fine-grained semantic recognition, event description, and reasoning, we introduce HyrCap for Semantic Action Understanding: calibrated proposals can be used for standard TAD detection and as inputs for LLM-based semantic understanding. We further define a staged evaluation protocol to separately measure event localization ability, segment-level semantic understanding, and their combined event understanding performance. Experiments on public benchmarks show that HyrCap achieves state-of-the-art performance across multiple metrics and provides a more fine-grained paradigm and evaluation protocol for query-based event segmentation and annotation.

Feature caching accelerates Diffusion Transformers by reusing intermediate computations across denoising timesteps, yet existing methods share a fundamental blind spot: they treat the standard reference denoising trajectory as the optimisation target, approximating its per-step states and thus inheriting its output as an implicit quality ceiling. We argue that this target is arbitrary --- what truly matters is the ideal result, not the intermediate path taken to reach it. We propose IdealCache, a training-free framework that abandons step-wise imitation of the reference trajectory. Instead, it anchors caching on a per-input ensemble of ideal trajectories constructed from candidate target results, aggregates their geometry into a single graph, and discovers optimal non-uniform schedules via a constrained shortest path on this graph. By decoupling from the reference process, IdealCache breaks the performance bottleneck of trajectory-approximation; our schedules can attain near-lossless fidelity in generation and even surpass on task-specific metrics (e.g., maintaining higher semantic consistency in image editing). Without retraining, IdealCache achieves $6.35\times$ speedup on Qwen-Image, $5.84\times$ on Qwen-Image-Edit, and $4.32\times$ on HunyuanVideo, with competitive or superior quality across tasks.

Modern LLM APIs often reveal only top-$K$ logit scores and censor the remaining vocabulary. We study the per-position distribution-recovery limits of this access model. For censoring threshold $\tau$, the compatible teacher distributions form an identified set whose total-variation diameter is exactly $U_K=(V{-}K)\exp(\tau)/(Z_A+(V{-}K)\exp(\tau)),$ where $Z_A$ is the observed partition function. For KL recovery, we give a computable binary-endpoint lower bound and an asymptotically matching small-ambiguity upper bound, with an extension to reference-aware attackers. Experiments on a Qwen3 math-reasoning teacher reveal a layered extraction hierarchy: on-task top-$K$ distillation recovers 12% of private capability, full-logit distillation recovers 56% despite 99% KL closure, and generation-based extraction recovers 96%. Top-$K$ censoring therefore limits per-position distribution recovery but does not by itself prevent capability extraction, separating fidelity from transfer in prompt-only logit distillation.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a promising approach to enhance Instruction Following (IF) capabilities of large language models (LLMs). However, RLVR for instruction following remains sample-inefficient and prone to reward hacking, where LLMs exploit verification shortcuts rather than fulfilling the core intent. To address these challenges, we frame RLVR for instruction following as an integrated environment that unifies dynamic task generation and robust training. We introduce Instruction Following Decorator (IFDecorator}, a framework coupling difficulty adaptation, intent alignment, and hack diagnostics. It features: (1) a cooperative-adversarial data flywheel that yields challenging yet solvable tasks for sample-efficient training; (2) IntentCheck, a gating module to mitigate reward hacking; and (3) TripWires, a proactive diagnostic tool for eliciting hacking behaviors and quantifying their prevalence. Extensive experiments show our Qwen2.5-32B-Instruct trained with IFDecorator using only 3,625 training examples achieves 87.43% on IFEval (outperforming GPT-4o) and improves FollowBench by 4.2%, while preserving general capabilities. Diagnostics confirm that our method effectively reduces reward hacking. The approach generalizes across model architectures and scales. We will release code and data for future research.

Recognizing human value preferences is crucial for building reliable AI agents in user-centric fields like psychological counseling and strategic negotiation. Existing benchmarks rely on explicit questionnaires, assuming (1) users are always cooperative and willing to self-report; (2) value preferences remain static. However, humans may conceal true preferences if they feel probed, and may adopt distinct value preferences depending on the context. Consequently, explicit queries are insufficient for predicting value preferences reliably. To address this, we introduce Implicit Value Probing (IVP), a benchmark simulating realistic scenarios where agents implicitly infer value preferences through strategic, multi-turn conversations. IVP grounds these interactions using a diverse set of Persona Agents, covering a wide spectrum of personalities and values. We evaluate mainstream LLMs on IVP, including Qwen3, DeepSeek, Doubao, and GPT-5, observing substantial performance gaps. We further develop Value-Prober-8B, a specialized model internalizing structured cognitive reasoning via distillation. It employs strategic questioning to elicit concealed value preferences. Experimental results demonstrate that Value-Prober-8B can effectively balance probing accuracy with social acceptability, achieving capabilities comparable to frontier models.

This paper studies the regret analysis for parallel Gaussian process (GP) bandit optimization. The known regret upper bounds for the widely used GP batched upper confidence bound and GP batched Thompson sampling (GP-BTS) suffer from a multiplicative factor with respect to the batch size $Q$. To avoid this degradation, existing analyses require a polynomial number of uncertainty sampling (US) for $Q$ at the beginning of optimization. However, this initial US phase is often ineffective in practice. This paper shows that the regret upper bound without the multiplicative factor on $Q$ can be achieved without the initial US phase, using GP-BTS as an example. Furthermore, we show much better regret upper bounds in the noiseless setting than in the noisy setting, as in the sequential GP bandit setting.

Effective Continual Video Instance Segmentation (CVIS) requires balancing catastrophic forgetting with the plasticity to learn new tasks. Existing methods typically adapt pre-trained VIS models using trainable visual prompts, but suffer from two key limitations: they overlook the distinct degradation of frame-level spatial knowledge (e.g., misclassification) and video-level temporal knowledge (e.g., trajectory disruption), and rely on rigid prompt retrieval that struggles under task ambiguity and distribution shifts. To address these issues, we propose STEAM, a CVIS framework with Spatial-Temporal Balanced Mixture-of-Experts Adapters. Instead of explicit retrieval, we introduce a Distribution-Aware spatial-temporal Scaling and sHifting (DASH) mechanism to align the pre-trained feature space with sequential tasks and reduce the domain gap. Furthermore, our method adaptively learns Task-Adaptive Spatial-Temporal mixture-of-Experts (TASTE) at both the frame and video levels, while updating them with Subspace-Constrained Orthogonal Adaptation (SCOA) to mitigate interference. Extensive experiments across diverse CVIS settings show that STEAM preserves temporal-spatial coherence and achieves state-of-the-art performance.

Flexible image tokenizers aim to represent an image using an ordered 1D variable-length token sequence. This flexible tokenization is typically achieved through nested dropout, where a portion of trailing tokens is randomly truncated during training, and the image is reconstructed using the remaining preceding sequence. However, this tail-truncation strategy inherently concentrates the image information in the early tokens, limiting the effectiveness of downstream AutoRegressive (AR) image generation as the token length increases. To overcome these limitations, we propose \textbf{ReTok}, a flexible tokenizer with \underline{Re}dundant \underline{Tok}en Padding and Hierarchical Semantic Regularization, designed to fully exploit all tokens for enhanced latent modeling. Specifically, we introduce \textbf{Redundant Token Padding} to activate tail tokens more frequently, thereby alleviating information over-concentration in early tokens. In addition, we apply \textbf{Hierarchical Semantic Regularization} to align the decoding features of earlier tokens with those from a pre-trained vision foundation model, while progressively reducing the regularization strength toward the tail to allow finer low-level detail reconstruction. Extensive experiments demonstrate the effectiveness of ReTok: on ImageNet 256$\times$256, our method achieves superior generation performance compared with both flexible and fixed-length tokenizers.

Neural processes (NPs) are flexible uncertainty-aware meta-learners, but often fail when meta-training tasks are scarce, losing generalization and collapsing into task memorization. We propose Context-Subset Self-Distillation Neural Processes (CSSDNP), a training framework that trains a student on random context subsets while distilling the full-context predictive distribution of an EMA teacher from the same task. This context-asymmetric objective combines reduced-context predictive learning with forward-KL distillation, enforcing conditional consistency across different context views without architectural or test-time changes. Theoretically, we show that context-subset training learns the reduced-context prediction problem while augmenting each task, and that calibrated full-context distillation preserves the asymptotic Bayes target while improving finite-sample estimation via teacher-guided shrinkage. Empirically, CSSDNP improves robustness and generalization across NP variants and tasks, especially in low-meta-data regimes.

Sparse dictionary learning (and, in particular, sparse autoencoders) attempts to learn a set of human-understandable concepts that can explain variation on an abstract space. A basic limitation of this approach is that it neither exploits nor represents the semantic relationships between the learned concepts. In this paper, we introduce a modified SAE architecture that explicitly models a semantic hierarchy of concepts. Application of this architecture to the internal representations of large language models shows both that semantic hierarchy can be learned, and that doing so improves both reconstruction and interpretability. Additionally, the architecture leads to significant improvements in computational efficiency.

Although existing deep multi-view clustering methods often achieve high performance, they predominantly focus on improving the fused representation while overlooking the enhancement of individual views, often resulting in stagnant or inconsistently improved view-specific representations that limit the final fusion quality. To address this, we propose Explicit single-View Enhancement (EVE), a novel deep multi-view clustering framework designed to explicitly optimize the clustering capability of each individual view. Specifically, EVE employs an attention-based module to generate a global consensus as a reliable anchor, which subsequently guides the refinement of individual views through two key components: feature-level selective alignment and structural-level neighborhood propagation. Crucially, the enhanced single views will further boost the fused representation, and finally leading to a mutually reinforcing closed-loop feedback process. Our method is also naturally applicable to incomplete deep multi-view clustering scenarios. Extensive experiments on multiple benchmark datasets demonstrate that EVE consistently outperforms state-of-the-art competitors, achieving substantial average improvements of 5.9% in ARI across all datasets.The code is available in the Supplementary Material.

Latent dynamical models can accurately fit neural population activity, yet accurate activity fitting alone does not guarantee mechanistic validity. Using synthetic benchmarks, we show that even well-fitted low-rank RNNs can yield misleading circuit interpretations when the prescribed activation function deviates from the ground-truth one. To address this limitation, we introduce gain-modulated linear dynamical systems (gmLDS), which decompose latent dynamics into a state-dependent, unit-wise gain and a static low-rank connectivity matrix, allowing the model to adapt to diverse nonlinear responses without assuming a fixed activation function. Across multiple synthetic benchmarks, gmLDS accurately recovers the local effective connectivity of the underlying model, thereby reconstructing the linearized dynamics along neural trajectories. Applied to neural recordings from perceptual and context-dependent decision-making tasks, gmLDS yields interpretable hypotheses about dynamical structure, including attractor structure in perceptual decision-making and context-dependent selection mechanisms. Together, these results support gmLDS as an effective approach for inferring computational structure from neural recordings.

Processing ultra-long multimodal inputs efficiently remains a critical bottleneck for Vision-Language Models (VLMs). This systematic challenge is essentially three-fold: (1) \textbf{Efficient Architecture}: compressing redundant visual sequences without sacrificing fine-grained details; (2) \textbf{Knowledge Transfer}: seamlessly migrating the capabilities of strong pretrained VLMs into more efficient novel architectures; and (3) \textbf{Generalization}: maintaining robust perception and reasoning across diverse multimodal domains. To address this, we introduce \textbf{InfiniteVL}, a linear-sparse hybrid VLM framework designed for efficient long-context understanding. At its core, InfiniteVL pairs a linear architecture to compress long-range visual context with sparse attention to retrieve precise local details. To support this architecture in moderate academic resources, we developed a three-stage knowledge transfer pipeline backed by capability-oriented data construction instead of training from scratch. This ensures smooth architectural alignment, rapid capability recovery, and robust adaptation across diverse domains. Furthermore, we adapt this foundation to specific deployment scenarios, deriving Sparse InfiniteVL for long-video analysis and Streaming InfiniteVL for real-time scene perception. Extensive experiments demonstrate that InfiniteVL matches the performance of leading VLMs while achieving a 1.7$\times$ decoding speedup and a 2.7$\times$ reduction in memory. Notably, Sparse InfiniteVL accelerates prefilling by 5$\times$ at a 256K context length, while Streaming InfiniteVL delivers stable 25 FPS real-time processing with a constant memory footprint.

Information-theoretic generalization bounds based on the supersample construction are a central tool for algorithm-dependent generalization analysis in the batch i.i.d.~setting. However, existing supersample conditional mutual information (CMI) bounds do not directly apply to sequential decision-making problems such as online learning, streaming active learning, and bandits, where data are revealed adaptively and the learner evolves along a causal trajectory. To address this limitation, we develop a sequential supersample framework that separates the learner filtration from a proof-side enlargement used for ghost-coordinate comparisons. Under a row-wise exchangeability assumption, the sequential generalization gap is controlled by sequential CMI, a sum of roundwise selector--loss information terms. We also establish a Bernstein-type refinement that yields faster rates under suitable variance conditions. The resulting bounds specialize to online learning, streaming active learning with importance weighting, and stochastic multi-armed bandits.

Running dense, open-weight large language models on commercially accessible workstations or single-GPU cloud setups is increasingly desirable due to cost and privacy constraints, particularly on tasks like coding and reasoning. Starting at around 24B parameters, however, they exceed the available GPU VRAM on these setups, forcing reliance on offloading methods that limit throughput, while leaving substantial CPU compute underutilized. Heterogeneous CPU–GPU speculative decoding is the dominant lossless-acceleration framework in this regime, but existing methods are limited by what we refer to as the **heterogeneity gap**: their CPU and GPU do not co-execute, their draft is a separate set of weights rather than a subnetwork of the verifier, and their VRAM footprint cannot scale to fill the available budget. We observe that channel-saliency methods induce an ordering on the transformer’s FFN channels so that the top-$k$ prefix closely approximates the full output, with graceful degradation as $k$ decreases. Building on this observation, we introduce **Speculative Tensor Parallelism (STeP)**, a training-free, self-speculative method that extracts a subnetwork of the verifier as a GPU-resident draft, places the remainder on the CPU, and verifies via concurrent CPU and GPU computation, provably preserving the verifier’s sampling distribution. Across dense models from 24B to 123B, GPU memory budgets from 32 to 192 GB, and five benchmarks, STeP outperforms SpecExec and SubSpec by up to $1.4 \times$ and $1.8 \times$ respectively, surpasses either tensor parallelism or speculative decoding alone, maximally utilizes the available VRAM, and closes the heterogeneity gap.

Text-to-CAD generation aims to make CAD modeling more accessible by generating executable CAD programs from natural-language descriptions. However, practical CAD workflows require more than executable geometry: generated programs should also expose reusable structure that supports later modification. We introduce intent-oriented text-to-CAD generation, where design intent is expressed as recoverable code-level signals: semantic features, reusable parameters, and lightweight parametric relations. Rather than changing the model architecture or decoding procedure, we investigate how supervision representation granularity affects both CAD generation and zero-shot intent-preserving editing. To enable this study, we construct IntentCAD-100K, a 100K paired text--code dataset built by converting low-level CAD construction traces into executable CadQuery programs. We further construct IntentCAD-Edit-1K to evaluate zero-shot intent-preserving editing from the same representation perspective. Intent2CAD uses a two-stage conservative lifting pipeline: semantic lifting rewrites reliable primitive groups into feature-level operations, while parametric lifting exposes lightweight parametric relations. Under controlled fine-tuning, semantic lifting mainly improves generation reliability and feature recovery, reducing IR from 2.39% to 0.22% and achieving 99.96% API Recovery, while parametric lifting makes reusable relation structure explicit and further improves relation-aware zero-shot editing.

As intelligent assistants are increasingly deployed in real-world environments, recommendations move from passive preference matching to grounding user intent from underspecified multimodal evidence. In these settings, users often provide only a reference image and a fuzzy natural-language query, leaving crucial visual attributes, personalized preferences, and situational constraints implicit. Existing recommenders either rely on historical interactions enriched with multimodal item features or presuppose fully specified textual requests, which are insufficient to resolve such queries in a single-turn setting. We present \textbf{IntentLens}, a tool-augmented recommendation framework that grounds underspecified multimodal queries into explicit, query-relevant evidence before ranking. IntentLens adopts a \emph{ground-then-rank} design: a shared multimodal LLM orchestrates visual, user-memory, and item-attribute tools to (i) recover latent user intent from the image and language query, and (ii) enrich each candidate with fine-grained, query-conditioned evidence; a lightweight ranker then scores candidates over the grounded representations. To support this new setting, we further construct two benchmarks with underspecified multimodal queries, \textsc{GoogleReview-MMTool} and \textsc{Yelp-MMTool}. Comprehensive experiments show that IntentLens outperforms strong MLLM and retrieval baselines.

Decision tree optimization is fundamental to interpretable machine learning, yet most tree learning algorithms are restricted to binary splits. This restriction can produce unnecessarily deep trees when the underlying decision logic is naturally multi-valued. Multiway-split trees address this, but finding sparse yet accurate ones is more challenging, as the number of possible partitions at each node grows combinatorially with the number of feature bins. We introduce SPLINTER (SParse Lookahead for Interpretable N-way Trees by Eliminative Ranking), a dynamic programming and branch-and-bound-based framework that combines efficient candidate generation with pruning bounds to find sparse, accurate multiway-split trees. Our empirical results show that our methods produce trees that achieve higher accuracy at a greater decision sparsity (shorter path lengths) than greedy multiway and near-optimal binary-split tree baselines, while remaining practical to train. We further extend the framework to approximate the Rashomon set of near-optimal multiway-split trees, allowing users to inspect multiple sparse and accurate alternatives.

Image-free zero-shot classifier expansion aims to extend recognition to unseen classes by establishing an association mapping between class semantics and visual classifier weights, without accessing original images. However, most existing methods rely on an association-based optimization mechanism, which struggles to meet the demands for fine-grained classification of visually and semantically similar classes. Since co-occurring attributes dominate class semantics, the model gravitates towards co-occurring attributes rather than discriminative attributes. This leads to geometrically entangled classifier weights, resulting in ambiguous decision boundaries. To address this problem, this paper proposes a novel approach that reframes weight generation through the lens of causal intervention. The method first employs comparative reasoning to disentangle discriminative and co-occurring attributes. Subsequently, it synthesizes intervened semantics to impose invariance and sensitivity constraints. This mechanism compels the model to suppress co-occurring attributes while enhancing sensitivity to discriminative attributes. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches, yielding clearer decision boundaries in fine-grained settings.

Muon and related norm-constrained matrix optimizers have become central to large-scale learning problems. They are formulated as a linear maximization oracle (LMO) over an ambient matrix-norm ball in unconstrained Euclidean space. However, these do not generalize cleanly to manifold-valued parameters such as low-rank factorizations, orthogonality constraints, or symmetric positive definite (SPD) matrices. Naively restricting the Muon LMO to the tangent space (i) breaks quotient symmetries and (ii) couples the tangent-space constraint with an ambient norm bound, thereby obstructing closed-form solutions on various manifolds of interest. We resolve both issues with a single observation: every Riemannian metric canonically lifts a unitarily invariant Euclidean norm to an intrinsic norm on each tangent space, and the resulting intrinsic norm constrained LMO is symmetry preserving. Building on this, we introduce intrinsic Muon (iMuon), a unified framework that yields closed-form updates on the fixed-rank, SPD, Stiefel, and Grassmann manifolds for any unitarily invariant norm, including the spectral, Frobenius, and nuclear norms. We establish convergence guarantees for both deterministic and stochastic iMuon with rate constants that depend only on the manifold dimension. Notably, on the fixed-rank manifold this constant depends only on the rank, making the rate independent of factor conditioning and removing the runtime factor-rescaling required by prior work. Experiments on LoRA finetuning of LLMs, image classification, and subspace learning illustrate the efficacy of the proposed approach.

Model-based and model-free reinforcement learning are traditionally viewed as separate paradigms: while the former learns an explicit model of the transition dynamics $P$, model-free agents typically estimate value functions tied to a specific policy and reward. In this paper, we challenge this dichotomy by proving that value-based agents trained on a sufficiently rich set of reward functions, e.g. using goal-conditioned RL, implicitly encode a unique and accurate world model. To extract this model in practice, we introduce $P$-learning; analogous to $Q$-learning, which approximates $Q^\star$ using samples from the environment, $P$-learning extracts an agent's model of the environment $P^\star$ by sampling from its $Q$-values, policies, and rewards, effectively inverting the Bellman equation. In deterministic MDPs, we prove that the true kernel $P$ can be recovered from an agent trained on a single generic goal for finite state spaces $\mathcal{S}$, and a finite number of Gaussian goals for continuous $\mathcal{S} \subseteq \mathbb{R}^d$, provided $Q$-values are accurate. In the stochastic setting, these conditions generalise to larger sets of goals depending on the reward function family. Even when our assumptions are violated, we empirically demonstrate that agents trained with a small number of sparse rewards encode accurate dynamics in (i) stochastic variants of FourRooms, (ii) MountainCar, and (iii) Reacher. This is further validated by training policies inside the extracted world model, for goals far beyond the training distribution, suggesting that goal-conditioned agents secretly contain implicit generalisation capabilities and providing a new lens into the connection between model-based, model-free, and goal-conditioned RL.

We introduce InvestigationWorlds, an agentic environment for legal investigation. We build on an underused artifact of U.S. civil litigation: the summary judgment motion. This motion relies upon a record composed of real evidence exhibits, and results in a court-adopted hypothesis that is treated as ground truth for the purposes of deciding the motion. Each environment is built from a real U.S. Federal Court case retrieved from Public Access to Court Electronic Records (PACER) and augmented by an attorney-validated generation pipeline that synthesizes role-tagged documents around the original record. The resulting corpus admits multiple coherent factual readings, only one of which matches the court-adopted hypothesis. Evaluating on 100 cases, we find agents often commit to incorrect hypotheses despite retrieving relevant evidence, struggling to distinguish the court-adopted hypothesis from alternative hypotheses.

Text watermarking helps identify AI-generated content, but its effect on factual reliability remains underexplored. In this paper, we study watermarking hallucination: factual errors induced or amplified by watermarking even when the required evidence is present in the context and the unwatermarked model can answer correctly. Using a controlled RAG setting, we compare paired unwatermarked and watermarked generations under the same condition. Across representative watermarking methods, including KGW, SWEET, DiPmark, GumbelSoft, Gumbel-Max, and SynthID-style watermarking, we find that watermarking hallucination is widespread: watermarked outputs can remain fluent while introducing factual errors. We attribute this failure mode to two mechanisms: (1) direct token-level bias, which can suppress fact-consistent tokens, and (2) prefix-induced drift, which accumulates through autoregressive decoding and weakens later attention to factual context. Motivated by this analysis, we propose Fact-Preserving Token Intervention (FPTI) and Fact-Preserving Attention Intervention (FPAI), two plug-in interventions that can be integrated into existing watermarking methods to improve factuality. Our experiments show that combining FPTI and FPAI mitigates around 90\% of watermark-induced hallucinations while preserving fluency and comparable decoding efficiency. Overall, this work highlights factuality as a first-class criterion in watermark evaluation, alongside detectability and robustness, and calls for careful factuality validation before deploying watermarks in fact-critical applications.

Real-world thermal infrared face reconstruction faces the dual challenges of compound sensor-specific degradations and a large infrared-to-visible domain gap. Existing methods either rely on oversimplified degradation assumptions or target only specific degradation types, leaving realistic sensor noise under cross-modal translation largely unaddressed. To tackle this compound problem, we decompose it into two sequential tasks, infrared degradation restoration to remove sensor-specific artifacts, followed by infrared-to-visible translation to bridge the domain gap to the visible spectrum. Both tasks require multi-scale detail reconstruction from spatially heterogeneous tokens that exhibit semantic, frequency, and task-specific variations. This motivates our proposed IRCasDiff, a cascaded diffusion framework consisting of the above two sequential stages. Both stages share a unified backbone with an asymmetric Sparse Mixture-of-Experts design, featuring content-adaptive routing in encoder blocks, frequency-decoupled reconstruction in decoder blocks, and text-free prompt generation for infrared-specific conditioning. The translation stage is initialized from converged restoration stage weights, enabling the model to master infrared restoration before tackling cross-modal mapping. Experiments on MCXFace and SpeakingFaces demonstrate state-of-the-art performance, with FID reductions up to 48.4\% over translation baselines and SSIM improvements up to 19.3\% over restoration baselines.

Reinforcement learning (RL) with verifiable rewards has driven recent gains in large language model (LLM) reasoning, but whether it creates new capabilities or merely sharpens existing ones remains debated, especially for compositional generalization, where models must combine learned primitives into novel multi-step functions. Using a controlled program synthesis domain over a typed domain-specific language (DSL) of 119 primitives with verifiable compositional depth and irreducibility, where the base model has zero capability prior to fine-tuning, we show on Qwen3-8B (and replicate within-family at 4B and 32B, and cross-family on Llama-3.1-8B) that compositional capability is established during supervised fine-tuning (SFT) and depends on irreducibility at two levels of SFT data: structural irreducibility of compositional exemplars and observational irreducibility, where prompt input-output pairs are chosen adversarially to rule out shallower programs. Neither alone is sufficient; even moderate contamination with structurally reducible exemplars substantially degrades performance, and full contamination yields zero pass@64. Given this SFT foundation, standard outcome-reward RL (GRPO, PPO) maintains but does not extend compositional performance, contrasting with prior results in settings with simpler nested function execution. Compositional capability is bottlenecked by structural properties of SFT data, before RL begins. Code release pending institutional IP review.

The linear representation hypothesis (LRH) holds that high-level concepts are encoded as linear directions in a model's feature space, and grounds much of mechanistic interpretability, from probing to sparse autoencoders to steering vectors. We argue the standard test for it is too lenient. A high linear-probe accuracy shows that a concept is linearly readable from the representation, not that it is one of the representation's privileged directions of variance, which is what the LRH actually claims. These two readings come apart in any high-dimensional, richly structured representation, and the gap matters. The structure that linear probing ignores is precisely where compositional generalization, out-of-distribution behavior, and the validity of post-hoc feature decomposition live. We propose three asymmetric diagnostics that operationalize the strong reading: reverse predictivity, spectral concentration, and out-of-distribution generalization. We validate them in a fully observed simulation where ground truth is known, then apply them to pretrained vision encoders on SVG-World, our deterministic generative benchmark with paired visual styles. Pretraining produces structure beyond what a forward probe can certify, but the linear-feature reading systematically overstates how much. Probing on its own is not enough to license the mechanistic claims that rest on it.

Why does the low dimensionality of representations, typically $d\approx 1000$, not prevent modern embedding-based retrieval models from scaling to billions, or even trillions, of data points? To answer this question, we study maximal-margin embeddings in the following retrieval model, classically studied in communication complexity [PS86] and more recently in embedding-based retrieval [WBNL25]. Let $A$ be a binary $N\times n$ matrix indicating whether each of $N$ queries is relevant to each of $n$ documents. We are interested in the largest margin $m>0,$ denoted by $\mathsf{m}^{\mathsf{rd}}(d, A),$ for which there exist unit norm embeddings of the queries and documents $U_1, U_2, \ldots, U_N, V_1, V_2, \ldots, V_n$ with the following property. $\langle U_j, V_i\rangle \ge m$ whenever $A_{ji} = 1$ and $\langle U_j, V_i\rangle \le -m$ when $A_{ji} = 0.$ A large margin is a key proxy for representation quality: it controls both robustness to perturbations and compositional generalization across queries. Our main theorem establishes that the best possible margin without a restriction on the dimension, $\mathsf{m}^{\mathsf{rd}}(+\infty, A),$ can be nearly achieved in dimension $d = O(\mathsf{m}^{\mathsf{rd}}(+\infty, A)^{-2}\log n)$ which improves a theorem of [BDES02]. Together with a matching lower bound in Theorem 1.5, we conclude that when $A$ is the $\binom{n}{k}\times n$ matrix containing all possible $k$-sparse rows once, dimension $d = O(k\log (n/k))$ is necessary and sufficient for the maximal possible margin $\mathsf{m}^{\mathsf{rd}}(+\infty, A) = \Theta(k^{-1/2})$ in this setting. This fully resolves the setup of [WBNL25]. We also give several constructions for large margins when $d = o(k\log (n/k)).$ Our proofs are based on connections between this problem and the literature on compressed sensing and the restricted isometry property. Finally, we empirically test the InfoNCE and sigmoid losses for producing large margin embeddings and demonstrate a clear advantage of the sigmoid loss.

Deep learning models are known for their ability to memorize training data. While memorization is often linked to risks such as privacy leakage and poor robustness, a highly influential claim by~\citet{feldman2020longtail} argues that memorization is actually \textit{necessary} for generalization. Their conclusion is based on the observation that removing points with high memorization scores reduces test accuracy. Upon closer inspection of their work, we uncover four critical flaws in the underlying methodology: \textbf{(1) sampling bias} in their approximation algorithm inflates memorization scores; \textbf{(2) high false positive rate} in their definition of memorization, leads to misclassification of non-memorized points as memorized; \textbf{(3) unprincipled thresholding} that resulting in an ill-posed problem; and \textbf{(4) data leakage} skews the test accuracy results. To address these limitations, we introduce a modifications for correctly identifying and evaluating memorization, including higher sampling rates, modifying the original memorization definition to reduce the false positive rates, proposing a method to identify a principled score threshold, and employing test datasets especially designed to avoid data leakage. Having accounted for these errors, our results show that, in contradiction to the original work, removing truly memorized points does not cause a drop in accuracy, and in most cases, improves test performance. These findings call into question the necessity of memorization in deep learning and highlight the importance of mitigating its risks.

Evaluating and mitigating a generative system's susceptibility to jailbreak attacks is critical to its safe deployment. Given the number of deployable systems, full per-configuration evaluation and optimization is impractical. In this paper, we formalize the behavioral geometry of a population of models that, by leveraging previously evaluated and defended models, supports both efficient susceptibility prediction and effective defense transfer across a population. We apply the framework to 79 models spanning 24 providers and to 100 system configurations of a single base model. Simple methods that use the behavioral geometry reach an AUPRC of $0.94$ for susceptibility detection with $\approx98\%$ fewer probes relative to a full evaluation. Using the behavioral geometry to select which model to transfer an optimized defense from outperforms same-provider assignment ($+2\%$, $p = 0.03$) at no additional probe cost, with a set of three models sufficient to cover the population. Results are robust to hyperparameter selection and judge.

Graph neural networks for bot and fraud detection operate on attributed graphs, where adversaries can jointly modify local edges and continuous node features. We study node-level certification under a localized mixed threat model with up to $d$ edge flips in the target message-passing neighborhood and an $\ell_2$ Frobenius feature budget $\epsilon$. We instantiate hybrid randomized smoothing for this graph setting and derive two graph-specific certificates: a sound Sequential NP composition baseline for the joint topology--continuous-feature threat model, and a Hybrid NP certificate whose homogeneous edge-smoothing structure reduces the adversarial adjacency search to an $O(d^2)$ optimization over edge-deletion and edge-addition counts. Across PubMed, Amazon Fraud, YelpChi, and MGTAB, the Hybrid NP certificate yields non-trivial certified regions under joint perturbations. We also construct adaptive joint attacks and show failure cases on PubMed, Amazon Fraud, MGTAB and YelpChi for topology-only certified nodes, demonstrating that topology-only guarantees can overestimate robustness when continuous feature perturbations are allowed. The results support joint discrete--continuous certification as the appropriate robustness notion for attributed graphs with continuous features.

Reconstructing the connectome of the brain is a central goal of modern neuroscience, but the required electron microscopy (EM) imaging is extremely time-consuming. EM image super-resolution (SR) accelerates this acquisition for downstream neuron segmentation, yet existing methods do not adequately couple SR with segmentation for EM: SR lacks segmentation guidance, and segmentation cannot adapt to super-resolved features. In this work, we propose a joint SR and segmentation framework specifically designed for EM images. In the super-resolution part, we introduce segmentation as an auxiliary task via a shared encoder with task-specific decoders to provide a semantic prior, improving reconstruction fidelity. In the segmentation part, we introduce a lightweight partial diffusion model that initializes denoising from a partially noised SR embedding, producing a structural prior that closely approximates real HR features and enhances segmentation accuracy on super-resolved EM images. Extensive experiments across multiple datasets and degradation settings demonstrate the superior performance of our framework.

Fairness in machine learning typically relies on the assumption of clean training data. However, in high-stakes applications like healthcare and criminal justice, privacy mechanisms and proxy variables frequently introduce simultaneous noise into both sensitive attributes and target labels. While existing methods address label or attribute noise in isolation, they fail under this "double noise" regime, often inadvertently amplifying bias. In this paper, we propose $\textbf{Jointly Robust Fairness (JRF)}$, a unified framework that guarantees fairness under simultaneous data corruption. JRF integrates a Forward Loss correction to handle label noise with a novel Matrix-Inverse Robust Loss (MIRL) that algebraically recovers the true demographic parity gap from noisy attribute observations. We provide rigorous theoretical guarantees for our estimator, including a finite-sample concentration bound demonstrating that the sample complexity scales quadratically with the inverse of the noise transition determinant. Furthermore, our framework naturally generalizes to both symmetric and highly asymmetric noise distributions. Extensive experiments on the Adult, Bank, and COMPAS datasets demonstrate that JRF reduces fairness violations by over 91\% in high-noise regimes ($\rho \ge 0.3$) compared to standard robust baselines, maintaining strict fairness without significant degradation in predictive utility.

Large language models (LLMs) are increasingly deployed in K–12 education, yet existing benchmarks such as C-Eval, CMMLU, GaokaoBench, and EduEval measure only whether a model can answer an exam question, i.e., factual recall. Effective educational AI further requires curriculum cognition: the structured understanding of how knowledge is organized, including prerequisite chains, concept taxonomies, experiment–concept links, and pedagogical sequencing. Curriculum cognition is neither probed by current benchmarks nor explicitly taught by current instruction-tuning data. To close this gap, we introduce K12-KGraph, a curriculum-aligned knowledge graph extracted from the official People's Education Press textbooks, covering mathematics, physics, chemistry, and biology across primary, middle, and high school, with seven node types (Concept, Skill, Experiment, Exercise, Section, Chapter, Book) and nine relation types spanning taxonomy, prerequisite, association, verification, assessment, location, and order. Building on this single graph, we derive two complementary resources: K12-Bench, a 23,640-question multi-select benchmark across five graph-derived task families (Ground, Prereq, Neighbor, Evidence, and Locate) that jointly probe curriculum cognition; and K12-Train, a KG-guided supervised fine-tuning corpus of ~2,300 QA pairs synthesized from node attributes and edge semantics. Experiments expose a clear gap and a clear remedy: on K12-Bench, even the strongest proprietary model (Gemini-3-Flash) reaches only 57% exact match and the best open-source model (Gemma-4-31B-IT) only 46%, with Prereq and Neighbor being the hardest; yet under a strictly matched 2,300-sample SFT budget on Qwen3-4B-Base and Llama3.1-8B-Base, K12-Train consistently outperforms equally sized subsets of eight mainstream instruction-tuning corpora (OpenHermes, Infinity, UltraChat, WizardLM, DataFlow, LMSYS, SmolTalk, Tulu-3) on both GaokaoBench and EduEval, showing that structural curriculum grounding is remarkably sample-efficient for educational SFT. We release the graph, benchmark, training data, and the full construction pipeline.

Effective inference is critical for interactive large language model (LLM) serving, where both time-to-first-token (TTFT) and end-to-end (E2E) latency shape user experience. Speculative decoding has emerged as a promising solution that reduces inference latency by employing lightweight drafting and parallel verification. However, it introduces a new scheduling issue, i.e., traditional shortest remaining processing time (SRPT) scheduling based on output length estimation fails to work due to its paradigm shift in inference structure. In this paper, we explore the preemption mechanism of speculative decoding-based LLM serving systems and show that — besides the output length, a request's remaining service time depends on draft acceptance and the number of verification rounds. Building on this insight, we develop a novel termination-aware scheduling framework, named \textsc{AlmostDone}, which first \textit{identifies requests that are almost done} within several future local decoding steps. It achieves this by introducing a lightweight predictor built from runtime draft-and-verify states, which helps approximate the SRPT-like scheduling better than existing approaches. Then, \textsc{AlmostDone} \textit{determines whether preemption is worthwhile} through a switching-aware online policy to limit excessive running-set alterations. Experiments on an open-source vLLM serving system show \textsc{AlmostDone} substantially surpasses the default schedulers, reducing mean and median TTFT by up to 3.83$\times$ and 3.39$\times$, and mean and median E2E request latency by up to 1.39$\times$ and 2.33$\times$. Against the oracle length-based preemption baseline TRAIL+, it still achieves up to $2.0\times$ lower mean TTFT and $1.3\times$ lower mean E2E request latency.

Dynamic convolution improves the representational flexibility of CNNs by generating input-adaptive kernels, yet existing methods either inflate parameters linearly with the number of base kernels or sacrifice inference throughput due to runtime kernel assembly. We revisit dynamic convolution from the angle of cross-layer redundancy: a CKA analysis across modern CNN families shows that within-stage convolutional kernels are highly correlated, suggesting that an entire stage can be reparameterized by a single shared kernel together with lightweight, layer-specific transformations. Building on this observation, we propose KernelDNA, a parameter-efficient framework in which multiple convolutional layers in a stage share one ''parent'' kernel, while layer-specific ''child'' kernels are produced by a \textit{decoupled} adapter that splits modulation into (i) an input-dependent dynamic channel gate and (ii) static spatial and filter modulations that can be pre-fused into the parent kernel before deployment. We provide a theoretical justification grounded in tensor approximation: when two kernels have CKA similarity $\ge 1 - \delta$, the multiplicative adapter incurs a Frobenius approximation error $O(\sqrt{\delta})$, and the implicit gradient consensus across child layers provably reduces the Rademacher complexity of the hypothesis class. Across ImageNet-1K and MS-COCO, KernelDNA achieves state-of-the-art accuracy-efficiency trade-offs over diverse backbones (ResNet18/50, MobileNetV2, ConvNeXt-Tiny), reducing parameters by $1.2$-$5\times$ versus dynamic convolution baselines while retaining $90$-$99$ % of the standard-convolution throughput---outperforming KernelWarehouse and FDConv across all backbones, and ODConv on $4$ of $5$ settings at a fraction of its parameter count.

We propose kernel-gradient drifting, a one-step generative modeling framework that replaces the fixed Euclidean displacement direction in drifting models with directions induced by the kernel itself. Standard drifting is attractive because it enables fast, high-quality generation without distilling a large pretrained diffusion model, but its theory is currently understood mainly for Gaussian kernels, where the drift coincides with smoothed score matching and is identifiable. Our gradient-based reformulation exposes this score-based structure for general kernels: the resulting drift is the score difference between kernel-smoothed data and model distributions, yielding identifiability for characteristic kernels and a smoothed-KL descent interpretation of the drifting dynamics. Since kernel gradients are intrinsic tangent vectors, the same construction extends naturally to Riemannian manifolds and to discrete data via the Fisher-Rao geometry of the probability simplex. Across spherical geospatial data, promoter DNA and molecule generation, kernel-gradient drifting enables state-of-the-art one-step generation beyond the Euclidean setting without distillation.

Activation steering provides a simple, training-free mechanism for controlling attributes of generative models (e.g., sentiment, style, helpfulness). However, standard approaches such as Difference-in-Means (DiM) apply a single input-independent steering vector across all activations, limiting expressivity and ignoring the local structure of the activation space. We propose Kernelized Activation Steering (KAS), a unifying framework that lifts activation steering into a reproducing kernel Hilbert space (RKHS). KAS formulates steering as an optimization problem expressed purely via kernel evaluations, yielding an implicit, activation-dependent steering score without constructing explicit feature maps. Unlike DiM, KAS induces locally adaptive steering: each activation is modified according to its relative position with respect to source and target reference sets, producing a nonlinear steering field over the representation space. Importantly, DiM is recovered as a special case under a linear kernel, offering a principled interpretation of steering, while richer kernels (e.g., RBF) enable geometry-aware interventions. Across standard activation steering tasks, including jailbreaking LLMs and image style control, KAS consistently outperforms existing methods.

Metric-induced discrete flow matching (MI-DFM) exploits token-latent geometry for discrete generation, but its practical use is limited by two issues: heuristic schedulers requiring hyperparameter search, and finite-step path-tracking error from its first-order continuous-time Markov chain (CTMC) solver. We address both issues. First, we derive a kinetic-optimal scheduler for prescribed scalar-parameterized probability paths, and instantiate it for MI-DFM as a training-free numerical schedule that traverses the path at constant Fisher–Rao speed. Second, we introduce a finite-step moment correction that adjusts the jump probability while preserving the CTMC jump destination distribution. We validate the resulting method, GibbsTTS, on codec-based zero-shot text-to-speech (TTS). Under controlled comparisons with a unified architecture and large-scale dataset, GibbsTTS achieves the best objective naturalness and is preferred in subjective evaluations over masked discrete generative baselines. Additionally, in comparison with the evaluated state-of-the-art TTS systems, GibbsTTS shows strong speaker similarity, achieving the highest similarity on three of four test sets and ranking second on the fourth. Demo and code are available in the supplementary materials.

In standard multi-head attention, each head computes its output independently, with cross-head information exchange limited to the final output projection. Talking-heads attention introduces explicit inter-head mixing on attention maps, but is incompatible with FlashAttention and incurs significant overhead. We propose knocking-heads attention (KHA), which couples heads at the parameter level via a shared, diagonally-initialized projection applied to query/key/value features before the scaled dot-product attention. The shared projection does not mix signals along the head axis; instead, it is jointly optimized by gradients from all heads, providing an implicit regularization on the head-specific projections. The diagonal initialization recovers MHA at step 0 and lets cross-head structure emerge gradually. We instantiate KHA in two complementary forms: a linear variant (KHA-Linear) whose shared projection can be absorbed into the original weights at inference, adding zero parameters or FLOPs at deployment; and a non-linear variant (KHA-MLP) that obtains stronger gains by introducing a small shared MLP, at modest inference cost. Training a 6.1B parameter MoE model (1.01B activated) on 1T tokens, we observe substantially fewer loss spikes, more uniform per-head activation norms, and a +1.26 average improvement across 19 downstream benchmarks. KHA is compatible with FlashAttention, KV-caching, and existing attention variants (MHA, MQA, GQA, GTA, MLA).

Vision foundation models are increasingly moving beyond 2D to volumetric domains such as 3D medical imaging, where unified pretraining across different imaging modalities (i.e. CT, MRI, and PET) could provide foundational models for diverse clinical tasks. However, training such models requires mixing heterogeneous imaging domains, and current mixture strategies remain largely heuristic. In this work, we observe that different medical imaging domains scale at variable rates during pretraining, and knowledge transfer between domains is strongly asymmetric: training on one domain can substantially improve another, but the reverse may be much weaker. Interestingly, both MAE reconstruction loss and cross-domain transfer follow predictable power-law trends with domain-specific behaviors. Motivated by these findings, we formulate data allocation as a scaling-law optimization problem. The derived allocations reveal an interpretable hub-and-island structure: highly transferable domains emerge as hubs that benefit many others and deserve strategic allocation, while isolated domains act as islands requiring direct investment. Empirically, transfer-aware allocation outperforms data-proportional sampling by up to 58\% and generalizes well to unseen budgets with r=0.989. Downstream validation on disease classification and organ/lesion segmentation further confirms that the derived transfer-aware mixtures provide stronger pretrained representations for clinical 3D medical imaging tasks.

As LLM-based agents increasingly browse the web on users' behalf, a natural question arises: can websites passively identify which underlying model powers an agent? Doing so would represent a significant security risk, enabling targeted attacks tailored to known model vulnerabilities. Across 14 frontier LLMs and four web environments spanning information retrieval and shopping tasks, we show that an agent's actions and interaction timings, captured via a passive JavaScript tracker, are sufficient to identify the underlying model with up to 96\% F1. We formalise this attack surface by demonstrating that classifiers trained on agent actions generalise across model sizes and families. We further show that strong classifiers can be trained from few interaction traces and that agent identity can be inferred early within an episode. Injecting randomised timing delays between actions substantially degrades classifier performance, but does not provide robust protection: a classifier retrained on delayed traces largely recovers performance. We release our harness and a labelled corpus of agent traces \href{https://anonymous.4open.science/r/known_actions-B31B}{here}.

Efficient long-context inference requires compressing the KV cache under strict memory budgets without degrading model outputs. We formulate this problem as a constrained optimization: minimize the total attention-output distortion subject to a global average bit-width budget, where each token’s KV entry can be assigned a bit-width ranging from 0 (eviction), through mixed precision, to 16 (no compression). The central challenge is to estimate each token’s contribution to attention-output distortion. We address this via a first-order perturbation analysis of the attention mechanism, which yields a closed-form decomposition of the distortion. Specifically, the compression cost of each token factorizes into (i) a token-specific risk score that captures the sensitivity of the attention output to perturbations of that entry, and (ii) a precision-dependent quantization error determined by the chosen bit-width. Building on this factorization, we solve the global allocation problem via Lagrangian relaxation, reducing it to independent per-token decisions that can be efficiently computed. Experiments on LongBench and RULER with LLaMA-3.1-8B-Instruct and Qwen3-8B show that our method achieves comparable accuracy to uniform 4-bit quantization while using an average of 2 bits per entry (2× memory reduction). At an average of 1 bit, our method maintains substantially higher accuracy than both uniform quantization and token eviction baselines, which deteriorate sharply in this regime. The advantage concentrates on retrieval-sensitive subtasks, where dispersed evidence makes uniform allocation and hard eviction especially costly.

Multi-modal object Re-Identification (ReID) aims to achieve robust all-weather perception by leveraging complementary information across heterogeneous modalities. However, existing methods still suffer from limitations in both feature fusion and feature alignment. In terms of feature fusion, most approaches assume equal importance across modalities and do not explicitly account for the quality differences among modalities in dynamic environments. As a result, the fused representations tend to be suboptimal and less discriminative in complex scenarios. In terms of feature alignment, existing methods typically rely on point-wise feature matching while ignoring relational structures among samples, which suppresses discriminative information and introduces cross-modal inconsistencies. To address these limitations, we propose LAMP, a unified framework that rethinks multi-modal interaction through language-modulated control and geometry-consistent alignment. We first adopt a Semantic-Assisted Feature Encoder (SAFE), which incorporates textual semantics into visual features to enhance local discriminative cues and provide more informative representations for subsequent processing. Building upon this, the Language-Modulated Dynamic Routing (LMDR) module generates modality-aware gating signals prior to feature fusion, enabling the adaptive adjustment of modality weights under challenging conditions. Furthermore, to address cross-modal inconsistencies, we introduce a Fused Geometric Prototype Alignment (FGPA) objective based on Fused Gromov-Wasserstein (FGW) optimal transport. Instead of enforcing strict point-wise alignment, FGPA establishes cross-modal consensus at the identity-level relational geometry, thereby preserving discriminative structures while mitigating inconsistencies across modalities. LMDR improves the stability of multi-modal fusion, while FGPA enhances structural consistency in cross-modal alignment. Together, they improve the discriminative capability of the model. Extensive experiments on three multi-modal object ReID benchmarks demonstrate the effectiveness of our method.

The ability of machine learning models to store input information in hidden layer embeddings, a form of model `memory', is widely employed but not well characterized. We find that causal language model embeddings typically contain relatively little input information regardless of data and compute scales during training. In contrast, embeddings from autoencoders trained for input regeneration are capable of nearly perfect memory formation. The substitution of memory embeddings for token sequences leads to computational efficiencies, motivating the introduction of a parallelizable encoder-decoder memory model architecture. Upon causal training these models contain information-poor embeddings incapable of arbitrary information access, but by combining causal and information retention objective functions they learn to form and decode information-rich memories. Training can be further streamlined by freezing a high fidelity encoder followed by a curriculum training approach where decoders first learn to process memories and then learn to additionally predict next tokens. We conclude that next token prediction training alone is poorly suited for accurate memory formation, motivating the use of combined objective functions for models where the entire input is not exposed.

Large reasoning models (LRMs) reach competition-level math and coding accuracy via long autoregressive decoding, making per-token decoding cost a primary deployment concern. Weight quantization is the standard tool for acceleration, but representative recipes --- including state-of-the-art end-to-end (E2E) QAT --- lose accuracy on long-decoding reasoning benchmarks despite preserving perplexity and short-decode accuracy. Through a systematic gradient-direction analysis, we identify two factors driving this gap: (i) KV-cache fidelity preservation under the QAT loss, which E2E supervision attenuates via the softmax Fisher metric; and (ii) Hessian-subspace alignment between calibration data and the deployment distribution. We propose LookAhead Quantization (LAQuant), a layer-wise weight-only QAT method that addresses both factors without online-transform overhead by combining reasoning-domain calibration with a one-layer lookahead loss whose implicit cross-layer co-adaptation preserves the next-layer residual stream. For Qwen3-4B under W3G128 quantization, LAQuant improves AIME25 Pass@1 over ParoQuant by 15.11pp (1.93pp over ParoQuant++ at matched calibration) while achieving a 3.42$\times$ decoding speedup over FP16 on RTX A6000, compared with ParoQuant's 3.01$\times$.

Behavior policies are often formulated as continuous generative models, whose iterative denoising processes are expressive but difficult to interpret and prone to producing implausible actions. We propose the Large Discrete Policy (LDiP), a fully discrete behavior modeling framework that selects actions from a large vocabulary of physically plausible candidates. Rather than perturbing actions, LDiP improves expressivity through stochastic iterative scoring: it progressively re-scores and prunes candidates with score-space stochasticity, enabling fine-grained ranking and exploration among plausible actions while preserving an explicit decision process. Across end-to-end planning, closed-loop driving, robotic manipulation, and vision-language-action settings, LDiP consistently outperforms strong discrete and continuous baselines in autonomous driving, and exceeds or matches continuous generative policies in robotic manipulation. These results show that discrete policies, when equipped with effective scoring mechanisms, offer an expressive, plausible, and interpretable alternative for behavior modeling.

Large language models (LLMs) trained on next-token prediction exhibit remarkable in-context learning (ICL) abilities, yet the representations that support ICL remain poorly understood. We consider such representations in a controlled setting: prompting LLMs with data emitted from hidden Markov models (HMMs) and probing for the corresponding belief state -- the posterior distribution over the HMM's hidden states given the observed token history. Across six open-source LLMs prompted with data from 40 HMMs selected for non-trivial belief structure, we find that belief states are linearly decodable from residual stream activations, with peak probe $R^2$-values ranging from 0.83 to 0.99 across HMM and LLM combinations, typically occurring at early-to-middle layers. To establish functional relevance, we intervene directly on the probe-identified subspace via patching and steering, resulting in downstream prediction quality on the order of the untampered model, while controls degrade substantially. Together, these results provide representation-level evidence that ICL in pre-trained LLMs approximates optimal Bayesian prediction over a context-inferred HMM. More broadly, our findings extend prior results linking input-distribution structure to activation geometry: from toy networks trained explicitly on HMM data to production-scale transformers trained on natural language.

Choosing a Large Language Model (LLM) for a given task requires comparing many strong candidates, yet standard evaluation relies on costly annotations over fixed evaluation sets. To address this challenge, we develop Select-LLM, the first framework for active model selection of LLMs. Select-LLM aims to find a small set of queries whose annotations are most informative for identifying the best LLM for a given task. To this end, we introduce a query selection rule based on expected information gain, computed from pairwise similarities between candidate model outputs. Because this rule only uses generated model responses, Select-LLM can be applied across candidate models without assumptions about their architecture or access to model weights. This makes it suitable for both open-weight and black-box LLMs. We evaluate Select-LLM across 23 datasets, 156 evaluated models, diverse task families, and multiple text evaluation metrics. Across all experiments, Select-LLM improves over the strongest baseline in every setting, with annotation cost reductions up to 81.8% for best model selection and up to 84.78% for near-best model selection.

Addressing missing modalities is an important challenge in multimodal image analysis and often relies on complex architectures that do not transfer easily to different datasets without architectural modifications or hyperparameter tuning. While most existing methods tackle this problem in feature space by engineering representations that are robust to missing inputs, we instead operate in weight space. We propose LARGO, a hypernetwork that compresses the $2^N-1$ dedicated missing-modality models into a single network by modelling the convolutional weights using the Canonical Polyadic (CP) tensor decomposition. Extensive experimental validation on BraTS 2018 (4 modalities, 15 scenarios) and ISLES 2022 (3 modalities, 7 scenarios) shows that our method ranks first in 47 out of 52 configurations, achieving average Dice improvements of +0.68\% and +2.53\% over state-of-the-art baselines (mmFormer, M³AE, ShaSpec, SimMLM). A proof-of-concept experiment on avMNIST suggests that LARGO may extend beyond medical imaging to heterogeneous non-medical modalities.

Retrieval-Augmented Generation (RAG) has become a standard approach for enhancing large language models (LLMs) with external knowledge, mitigating hallucinations, and improving factuality. However, existing systems rely on generating natural language queries at each hop and maintaining a strict architectural separation between retriever and generator, preventing them from leveraging the full representational capacity of the LLM. We propose LAnR (Latent Abstraction for RAG), a unified framework in which a single LLM jointly performs encoding, retrieval, and generation entirely within its own latent space. Rather than generating textual queries, LAnR produces dense retrieval vectors from the hidden states of a designated [PRED] token and uses them to match against encoded document representations from the same model. Furthermore, LAnR adaptively decides when sufficient evidence has been retrieved using a lightweight MLP control head over those same hidden states, eliminating explicit token-level stopping reasoning. Extensive experiments on five QA benchmarks spanning single-hop and multi-hop settings demonstrate that LAnR outperforms existing RAG methods, while achieving improved inference efficiency through reduced number of retrieval calls and tighter model integration.

Bridging independent vision-language models (VLMs) and small language models (SLMs) in a modular multi-agent system typically forces them to exchange information as discrete text, which creates a lossy bottleneck for spatial and fine-grained details. This constraint is especially acute for SLMs, whose limited capacity makes them more reliant on the quality of the input signal. While recent works on latent multi-agent communication have explored latent-space exchange between identical or homogeneous LLMs, propagating rich visual information across model boundaries without text generation has not been studied yet. We propose Latent-Lens, a lightweight codec that bridges the representation gap between a frozen VLM and a frozen SLM directly in latent space, enabling them to communicate purely through hidden states. The codec consists of a domain encoder that projects hidden states from the VLM into the SLM's embedding space, and an attention decoder formalised as a LoRA that teaches the SLM to attend to these visual prefixes, while training only 1.6M parameters (0.2\% of the combined model size). On free-form visual question answering (CLEVR, GQA), Latent-Lens achieves +10.2-28.3pp over a text-communication baseline given the same LoRA budget (+36.4-78.4pp when both agents are frozen); on image captioning (Flickr8k) it improves over the same baseline by +18.5 (+18.9 frozen). Replacing the SLM (TinyLlama-1.1B) or the VLM (InternVL2-1B) with alternative models of different architectures, vocabularies, and hidden dimensions yields consistent gains, confirming generality across heterogeneous model families.

Autoregressive (AR) models remain the dominant paradigm for natural language generation, but their strictly sequential decoding process leads to high inference latency. Recent diffusion-inspired language models, such as LLaDA and Dream, alleviate this issue through parallel generation, yet they still face two key limitations: information loss, since predictive distributions over non-finalised tokens are discarded at each step, and unstable commitment dynamics, where local token decisions are not sufficiently coordinated at the global level. We propose Latent Refinement Decoding (LRD), a two-stage decoding framework consisting of Latent Refinement and a Predictive Feedback Loop. In the first stage, LRD keeps masked positions as distributional mixtures of predicted tokens and the mask embedding, enabling the model to form more globally consistent beliefs before committing tokens. In the second stage, LRD progressively finalises confident tokens while preserving uncertain ones for further iterative refinement, with KL-divergence dynamics serving as a stable criterion for convergence and early stopping. Experiments show that LRD improves performance on both coding tasks, including HumanEval by 6.3 and MBPP by 2.6, and reasoning tasks, including GSM8K by 2.9 and MATH500 by 3.8, while achieving up to 10.6× overall speedup. Moreover, LRD is compatible with single-token decoding, multi-token threshold commitment, and system-level accelerators; when combined with Fast-dLLM, it reaches up to 18.4× speedup over vanilla decoding while further improving accuracy.

Random Feature Models (RFMs) are attractive PDE surrogates because their linear-in-features form reduces training to regularized least squares. Standard RFMs, however, use one stationary spectrum over the whole domain, creating a bandwidth dilemma on heterogeneous PDE solutions: fine scales resolve sharp local gradients but oscillate in smooth regions, while coarse scales miss interface-dominated derivatives. We propose \textbf{LDD-RFM}, a learnable domain-decomposition framework whose local experts remain linear RFMs and whose nonlinear variables specify compact-support partition-of-unity (PoU) gates and local spectral scales. For fixed gates and spectra, expert coefficients are eliminated by ridge regression through variable projection, yielding a differentiable reduced objective with explicit linear-solve and partition-smoothness controls. We prove an $H^1$ error decomposition in which PoU sharpness enters through a product-rule term. Controlled regression and PDE experiments show that LDD-RFM reduces $H^1$ errors by $10\times$ over global and fixed-partition RFM baselines under matched feature budgets; exact-solver, timing, and robustness studies check that these gains are not artifacts of the linear solver or evaluation protocol.

We present Leak-CURBER, the largest multimodal, leakage-controlled benchmark for evaluating machine learning and deep learning models for enzymatic reactions. Leak-CURBER standardizes data from 10 biochemical databases and curates datasets with multimodal inputs, metrics, and leakage-controlled splits for 4 tasks. The benchmark comprises 10 subtasks: prediction of reaction outcomes (2), enzyme function (2), kinetic parameters (3), and protein–ligand binding (3). Leak-CURBER establishes new evaluation protocols that control for train/validation/test overlap across protein sequence and structure, molecule string and structure, and reaction similarity, thereby exposing generalization failures of learning methods. Evaluating the state-of-the-art (SOTA) models on Leak-CURBER shows that their strong performance degrades sharply under these splits, often approaching random classification or retrieval performance, and near-zero regression performance. Leak-CURBER includes precomputed embeddings, continual benchmark releases, reproducible curation scripts, and a public leaderboard to support reliable comparison of enzymatic-reaction models.

Link prediction relies on pairwise structural proximity, but standard node representations are not explicitly designed so that their interactions reflect such signals. Existing structure-aware approaches address this gap through specialized architectures or handcrafted proximity features, often at the expense of generality or additional computational cost. In this work, we take a data-centric approach and propose $\text{LinkDPE}$, a learnable $\underline{\text{d}}$iffusion-based $\underline{\text{p}}$ositional $\underline{\text{e}}$ncoding framework that injects structural proximity into the node features. Our key idea is that diffusion kernels, which capture multi-scale connectivity patterns, admit a factorization into node-wise embeddings whose inner products recover pairwise proximity scores. Based only on the observed training graph, $\text{LinkDPE}$ constructs diffusion-based encodings across scales, adaptively selects complementary diffusion scales, and learns task-specific spectral weightings, enabling a wide range of link predictors to exploit structural information without architectural modification. Theoretically, we show that (i) diffusion proximity admits a node-wise kernel factorization, (ii) many classical link prediction heuristics are recovered as special cases of low-order walk and diffusion operators, and (iii) multi-scale spectral responses allow the model to capture a richer set of structural similarity patterns within a shared truncated spectral basis. Across 8 benchmark datasets and 5 representative base models, $\text{LinkDPE}$ consistently improves link prediction performance over original features and positional-encoding baselines, yielding average gains of 6.97\% over the unaugmented base models.

Implicit neural representations (INRs) are shaped by the spectral structure induced by their input encodings and activation functions. Existing methods improve fitting primarily by modifying which frequencies are available to the network, through coordinate encodings or periodic nonlinearities. However, frequency access is not the only bottleneck: signals with localized or spatially varying structure require the network to efficiently compose frequencies into multi-harmonic internal responses. We introduce learnable spectral activations (LSA), which replace fixed neuron-level nonlinearities with a residual truncated Fourier series whose harmonic amplitudes are learned during training. LSA does not expand the asymptotic function class. Instead, it changes the factorization of the representation: linear weights select features while activation coefficients control spectral shaping, and the two are updated by separate gradients. Because the activation coefficients enter the loss linearly given fixed pre-activations, spectral tuning becomes a simpler subproblem than in architectures where it is entangled with feature selection. Empirically, this factorization concentrates more target-signal energy in the leading eigenmodes of the neural tangent kernel, consistent with improved optimization behavior. Across audio, image, neural radiance field, and neural acoustic field tasks, LSA also improves reconstruction quality.

Generative models, from diffusion models to large language models, achieve remarkable performance but at a cost in training data orders of magnitude larger than what biological learners require. An alternative paradigm has emerged in which networks are trained to predict their \emph{own} latent representations of related views or masked regions, as in data2vec and JEPA -- an idea related to predictive-coding accounts of the cortex. Despite strong empirical results, the theoretical understanding of these methods remains limited. Central questions include: by how much does latent prediction actually improve data efficiency? Is there a benefit to stacking such methods into multi-scale hierarchies? We answer both using as data a tractable probabilistic context-free grammar that captures the compositional structure of natural language and images. We prove that latent prediction recovers the full latent tree of depth $L$ from a number of samples scaling as $m^3$, where $m$ is the number of production rules per symbol. This is much fewer than the $m^L$ samples required by supervised learning and the $m^{L+1}$ required by token-level SSL. We confirm this bound with (i) a hierarchical clustering algorithm, (ii) an end-to-end neural network whose predictor-clusterer modules predict their own latents at each level via gradient descent, and (iii) the first sample-complexity analysis of data2vec, which we show implicitly performs hierarchical latent prediction. This suggests that explicit stacking such as H-JEPA is largely redundant.

Recent progress in integrating vision-language-action (VLA) models with world modeling has significantly advanced end-to-end autonomous driving by allowing systems to reason about imagined futures instead of merely reacting to current observations. However, most world VLA models generate future frames at the pixel level with plenty of dense visual tokens. Although such generated futures can be visually plausible, they preserve many driving-irrelevant appearance details, i.e., only a small portion of the tokens directly encode planning-relevant factors, such as agent geometry, motion dynamics, and interactions. This introduces a significant performance bottleneck. An important question thus arises: what world state should a world VLA model reason over to support planning beyond pixel-level prediction? We argue that planning-oriented world modeling should focus on an agentic latent state that captures the interactions between objects in the driving scene over time. As such, we introduce AgentWorld, an agentic world VLA model that uses discrete agent tokens jointly for reasoning and trajectory planning. Each token is dynamically grounded to a surrounding agent associated with its object's state, such as semantic identity, 3D geometry, motion, and interaction context. AgentWorld learns these tokens through three key processes: explicit geometric grounding, reasoning-oriented supervised fine-tuning, and agent-oriented reinforcement learning. Extensive open-loop and closed-loop experiments under diverse driving scenarios show that AgentWorld enhances planning capabilities, fully exploits visual evidence and agent dynamics, and yields interpretable agentic world states. The code will be made publicly available.

The field of learning-augmented algorithms has demonstrated that machine-learned predictions can bypass worst-case lower bounds across a wide range of problems. So far, however, the focus has been almost exclusively on polynomial-time algorithms, where predictions improve competitive ratios, approximation guarantees, or running times. In this paper, we raise the question of whether predictions can push the frontier of exact exponential-time algorithms for NP-hard problems. We answer this question affirmatively by proposing a general approach that augments an entire family of state-of-the-art exact algorithms for a variety of subset selection problems. We show that a noisy predictor that is only marginally better than random guessing suffices to provably reduce the search space, and that the resulting runtime speedup scales smoothly with the prediction quality. Importantly, our algorithms require only pairwise independence of predictions or, alternatively, do not require the knowledge of the predictor's accuracy–both strictly weaker and more realistic settings than typically assumed.

In-context learning for tabular data sets strong predictive standards in observational settings; it however primarily relies on correlational structure, which becomes unreliable under distribution shift or intervention. While established methods to discover causal structure exist, they are often focused on structure identifiability and decoupled from the predictive architectures that could benefit from them. To bridge these perspectives, we study how to simultaneously infer and enforce causal structure in the form of topological variable orderings into tabular prediction. Unlike standard architectures, our model TabOrder uses causal order-constrained attention, basing predictions only on features that precede a target under a learned causal order. Similar to causal discovery methods, TabOrder learns the optimal variable ordering in an unsupervised manner through a likelihood-based objective. We justify this choice under standard functional model classes and also study how sample missingness, a common challenge in tabular data, interacts with causal direction identification. Empirically, we confirm that TabOrder recovers accurate variable orderings while addressing prediction and imputation tasks, as well as gives insight into real-world biological data under intervention.

Physical fields on meshes require a separation between topology and geometry: conservation laws are topological and should be exact, while geometry, material response, and anisotropic coupling must be learned from data. Existing neural surrogates often mix these roles inside unconstrained message passing. We introduce Riemannian Hodge Message Passing (RHMP), which turns this separation into an architectural principle. RHMP fixes the cellular coboundaries ($d_k$) determined by oriented incidence and learns symmetric positive-definite cochain metrics ($H_k$) for geometry-dependent propagation. Treating $H_k$ as the learned metric motivates cochain-frame equivariance: physical propagation should be invariant to orthogonal changes of the hidden cochain feature basis. RHMP implements this principle with metric-weighted Hodge blocks ($d_k^\top H_{k+1}d_k$), yielding exact cochain-complex identities ($d_{k+1}d_k=0$), nonnegative Hodge energies, positive-semidefinite operators, and exact Abelian curvature invariance. Across seven physical benchmarks spanning fluids, electromagnetism, gauge fields, and variable-mesh CFD, RHMP achieves the best overall performance, with the largest gains when topology, learned geometry, and field structure interact.

Vision-Language Pre-trained (VLP) models such as CLIP learn strong representations from large-scale image–text pairs and demonstrate impressive zero-shot transfer. However, by learning to align visual and textual features only at a global level, they often suffer from limited interpretability and weak fine-grained perception. This issue stems from pre-training data that overlooks key visual details of object parts, which are often important for distinguishing between subordinate categories (e.g., species of birds or models of cars). Multimodal Large Language Models (MLLMs) built on CLIP-style vision encoders inherit this weakness, limiting both accuracy and trustworthiness. To address this, we propose Part-Aware CLIP (PA-CLIP), a framework that improves fine-grained perception while enhancing interpretability. First, we leverage MLLMs to construct a new dataset,FG-Part, containing about one million part-level image–text pairs that explicitly describe discriminative components (e.g., beak shape or wing patterns). Second, we introduce a part-aware training strategy that encourages explicit grounding of fine-grained textual descriptions to corresponding image regions, strengthening part-level cross-modal alignment. Extensive experiments show that PA-CLIP achieves state-of-the-art results on multiple fine-grained visual recognition benchmarks, validating the benefit of part-level captions for capturing subtle details. Moreover, evaluations on general tasks such as cross-modal retrieval indicate that these improvements do not compromise the model's core generalist capabilities.

We present a system for learning generalizable hand-object tracking controllers purely from synthetic data, without requiring any human demonstrations. Our approach makes two key contributions: (1) HOP, a Hand-Object Planner, which can synthesize diverse hand-object trajectories; and (2) HOT, a Hand-Object Tracker that bridges synthetic-to-physical transfer through reinforcement learning and interaction imitation learning, delivering a generalizable controller conditioned on target hand-object states. Our method extends to diverse object shapes and hand morphologies. Through extensive evaluations, we show that our approach enables dexterous hands to track challenging, long-horizon sequences including object re-arrangement and agile in-hand reorientation. These results represent a significant step toward scalable foundation controllers for manipulation that can learn entirely from synthetic data, breaking the data bottleneck that has long constrained progress in dexterous manipulation.

Human videos contain rich manipulation priors, but using them for robot learning remains difficult because raw observations entangle scene understanding, human motion, and embodiment-specific action. We introduce MoT-HRA, a hierarchical vision-language-action framework that learns human-intention priors from large-scale human demonstrations. We first curate HA-2.2M, a 2.2M-episode action-language dataset reconstructed from heterogeneous human videos through hand-centric filtering, spatial reconstruction, temporal segmentation, and language alignment. On top of this dataset, MoT-HRA factorizes manipulation into three coupled experts: a vision-language expert predicts an embodiment-agnostic 3D trajectory, an intention expert models MANO-style hand motion as a latent human-motion prior, and a fine expert maps the intention-aware representation to robot action chunks. A shared-attention trunk and read-only key-value transfer allow downstream control to use human priors while limiting interference with upstream representations. Experiments on hand motion generation, simulated manipulation, and real-world robot tasks show that MoT-HRA improves motion plausibility and robust control under distribution shift.

Domain generalization aims to learn models that generalize to unseen target domains using labeled data from multiple source domains. Many existing approaches focus on learning domain-invariant representations, but enforcing invariance alone may fail to preserve sufficient task-relevant information under domain shift. We propose BLENDER, a reconstruction-aware framework for domain generalization that employs a hierarchical latent decomposition to separate domain-invariant and domain-specific factors. This structure prioritizes informativeness in the domain-invariant representation and captures residual domain-specific variation through conditional modeling. We further provide a theoretical analysis establishing a bound on the reconstruction risk for unseen target domains, revealing how domain invariance and latent disentanglement contribute to generalization beyond the observed domains. Experiments on standard domain generalization datasets demonstrate strong performance under substantial domain shifts, and qualitative analyses show that BLENDER promotes a structured allocation of information between domain-invariant and domain-specific latent representations.

We study learning rate (LR) scaling for hybrid architectures combining Transformer and State-Space Model (SSM) blocks, a class adopted by several recent production language models. In particular, we focus on the gap between the theoretical scaling rules derived for SSMs under zero-order-hold (ZOH) discretization at infinite width with proportional state size, and the field-standard practical implementations using simplified-ZOH Mamba at fixed state size. Surprisingly, in this practical regime hybrid architectures achieve a near-zero LR transfer gap across widths 256-2048 and depths 4-32 at sub-billion-parameter scale using only the original $\mu$P prescription, even though SSM operations fall outside its Tensor Programs representability conditions and every parameterization we test fails the standard coordinate-check diagnostic of $\mu$P correctness. We attribute this to a two-condition decomposition of LR transfer in hybrid architectures: a global update-to-weight invariance, enforced by $\mu$P's initialization and LR scaling; and a local per-component balance, provided by AdamW's $g/\sqrt{v}$ per-parameter normalization. Our observations show that the optimal LR is invariant to width up to $8\times$, as well as to depth, sequence length, batch size, and Transformer-to-SSM ratio, and transfers to Nemotron-H, a production hybrid outside our custom architecture set. We hope these findings fill the gap between theoretical scaling rules and practical hybrid implementations, and stimulate further research toward bridging it.

Learning in neural systems arises from synaptic changes that reshape the representations underlying behavior. While low-rank recurrent neural networks (RNNs) have emerged as a powerful framework for linking connectivity to function, a theoretical understanding of their learning process remains elusive. Here, we extend the low-rank framework from activity to learning by deriving gradient-descent dynamics directly in a reduced overlap space. We formulate a closed-form, low-dimensional system of ODEs that governs learning in this space, exact for linear RNNs and asymptotically exact for nonlinear RNNs in the large-$N$ Gaussian limit. Central to our analysis is a distinction between two classes of overlaps: _loss-visible_ overlaps, which fully determine network activity, output, and loss, and _loss-invisible_ overlaps, which do not affect function but are required to describe learning. We illustrate the consequences of this decomposition through two phenomena. First, we show that learning can serve as a perturbation that exposes differences in connectivity between functionally equivalent networks. Second, we show that loss-invisible overlaps can act as memory variables that encode training history, and characterize the conditions under which this occurs. Finally, we present several testable predictions for biological learning experiments derived from our theory.

Two-Stage Robust Optimization (2RO) with discrete uncertainty is challenging, often rendering exact solutions prohibitive. Scenario reduction alleviates this issue by selecting a small, representative subset of scenarios to enable tractable computation. However, existing methods are largely *problem-agnostic*, operating solely on the uncertainty set without consulting the feasible region or recourse structure. In this paper, we introduce PRISE, a *problem-driven* sequential lookahead heuristic that constructs reduced scenario sets by evaluating the marginal impact of each scenario. While PRISE yields high-quality scenario subsets, each selection step requires solving multiple subproblems, making it computationally expensive at scale. To address this, we propose NeurPRISE, a neural surrogate model built on a GNN-Transformer backbone that encodes the per-scenario structure via graph convolution and captures cross-scenario interactions through attention. NeurPRISE is trained via imitation learning with a gain-aware ranking objective, which distills marginal gain information from PRISE into a learned scoring function for scenario ranking and selection. Extensive results on three 2RO problems show that NeurPRISE consistently achieves competitive regret relative to comprehensive methods, maintains strong scalability with varying numbers of scenarios, and delivers $7$--$200\times$ speedup over PRISE. NeurPRISE also exhibits strong zero-shot generalization, effectively handling instances with larger problem scales (up to $5\times$), more scenarios (up to $4\times$), and distribution shifts.

Human motion prediction (HMP) aims to forecast future movement from an observed motion sequence, and recent methods have incorporated 3D scene geometry and object semantics to improve long-term plausibility. However, the destination of future motion remains ambiguous in cluttered scenes: the same observed motion can lead to different interaction targets or navigation modes. Existing methods either handle this ambiguity implicitly or rely on auxiliary signals like gaze that are hard to obtain in practice. To address this, we propose SGIP, which explicitly predicts where in the scene future motion will be grounded across both objects and open navigation regions, and uses it to condition future trajectory and full-body pose decoder. This scene-grounded interaction prior is learned via intent distillation: at training time only, a teacher model scores candidate scene regions by combining an LLM prompted with a textual description of the person's intent with geometric motion-scene features. A student model learns to reproduce these scores from only the observed motion and 3D scene, making intent and LLM unnecessary at test time. Across multiple benchmark datasets, SGIP outperforms prior scene-aware baselines on both trajectory and pose accuracy in a realistic deployment setting where neither gaze nor intent is available at inference, demonstrating effectiveness of the proposed framework.

State-of-the-art machine learning weather prediction systems achieve strong upper-air forecast skill but degrade at longer lead times, with mesoscale structures progressively vanishing and ensemble forecasts becoming under-dispersive. While these limitations are often attributed to the training objective, we find that message-passing processors significantly contribute to this degradation. Repeated neighborhood aggregation at each autoregressive step acts as a smoothing operator that, over rollout steps, suppresses high-frequency information in latent space. This effect is particularly limiting in regional weather prediction, where fine-scale structures are already partially unresolved, further eroding mesoscale variability and hindering long-horizon forecasting. We propose GraphMet, a graph-based weather model motivated by Koopman theory that replaces stacked message-passing layers with a single linear Koopman operator step. GraphMet learns linear evolution in a spectral latent basis while preserving the encoder, decoder, and multi-scale icosahedral mesh structure. A two-term spectral objective jointly learns a stable Koopman basis and preserves latent geometry, enabling efficient long-range propagation without repeated nonlinear aggregation. Empirically, GraphMet reduces the 15-day activity and improves regional RMSE by 11-14% at 5km resolution, and enables efficient probabilistic forecasting with lower inference cost than state-of-the-art methods. Our code available at : https://anonymous.4open.science/w/GraphMet-54ED/.

Survival analysis provides statistical methods to model the time until an event occurs. Reporting delays arise when event times are not observed at their occurrence but are only revealed upon reporting. This issue is particularly critical for timely risk evaluation when the observation window is short due to administrative censoring. In this study, we incorporate right-censored reporting delays by jointly modeling parametric hazards for the event and reporting processes. We then construct a consistent estimator for the model parameters and develop a Monte Carlo expectation-maximization algorithm to compute it. To address the challenges posed by administrative censoring, we leverage these findings and propose a transfer-learning procedure. Experimental results demonstrate that our method improves the accuracy of timely risk evaluation under administrative censoring.

When auditing machine learning models, human experts sequentially inspect model predictions and explanations, gradually building a mental model of the model’s behavior to uncover errors. This manual process is critical for safe deployment in high-stakes domains but fundamentally limited in scalability. In this work, we propose a novel method to simulate this human auditing behavior for automating the auditing process. We formalize auditing as a sequential decision-making problem, modeled as a Markov decision process. Our state representation summarizes the auditor’s accumulated knowledge of model predictions and feature usage. Building on this formalization, we solve the MDP using reinforcement learning, yielding RLAuditor, which learns to select informative test samples to efficiently uncover model errors. We further connect our formalization to Theory of Mind, showing that our state mirrors how human auditors build mental models: its update rule is equivalent to Bayesian belief revision and it converges to the model’s behavioral pattern at an interpretable rate. We validate our approach on multiple ML models and datasets across different data modalities. In a human subject study, we demonstrate that RLAuditor helps human auditors produce more accurate audit reports in less time.

Reinforcement learning (RL) post-training has become an important paradigm for improving multimodal large language models (MLLMs), especially with automatically verifiable rewards. For vision-centric MLLMs, self-supervised visual pretext tasks provide such rewards without human annotations. However, existing self-supervised RL methods mostly exploit within-instance structure, such as spatial or temporal ordering within a single image or video, offering limited supervision for cross-instance discrimination and fine-grained comparison. We introduce Group-and-Order Self-Supervised Reinforcement Learning (GO-SSL), a cross-instance jigsaw framework for vision-centric MLLM post-training. Given two visual instances, GO-SSL mixes their local elements and trains the model to group them by instance of origin while recovering the order within each group, coupling instance-level comparison with spatial, temporal, or geometric structure recovery. This paired formulation turns the same unlabeled data into richer verifiable supervision through diverse pairings, shuffles, and hard-pair curricula. We instantiate GO-SSL mainly on cross-image jigsaw tasks and further extend it to 3D depth and video temporal ordering. Across diverse benchmarks, GO-SSL consistently improves over the base MLLM and prior baseline methods, with clear gains in fine-grained perception and spatial understanding. These results suggest a data-efficient direction for self-supervised RL in MLLMs: constructing comparative cross-instance contexts can provide richer transferable supervision than merely scaling data volume or adding isolated pretext tasks. Code will be released later.

Autoregressive video generation has achieved impressive results on short clips, yet generating long videos remains challenging due to error accumulation over extended horizons, where each predicted frame depends on potentially imperfect previous frames. Existing approaches typically rely on static context management strategies, such as sliding-window caches that discard early frames or fixed sink frames that permanently preserve initial content. These designs either lose critical long-range context or anchor generation to outdated information, which can lead to degraded fidelity, motion stagnation, and reduced diversity. In this paper, we propose \textbf{Learning to Memorize (L2M)}, a learned context memory management framework that dynamically retains, evicts, and evolves historical frames according to predicted importance. The framework consists of an \emph{importance prediction router} that estimates the relevance of each historical frame, \emph{degradation-aware training} that improves robustness to noisy or corrupted context, and a \emph{dynamic memory initialization and evolution} mechanism that updates long-term memory using exponential moving average (EMA) importance scores. This allows high-quality frames to replace stale anchors and enables memory to adapt naturally to scene evolution. Trained with a teacher-student distillation objective and sparsity regularization, L2M makes efficient use of a fixed KV-cache budget while preserving the most informative contextual information. Extensive experiments show that L2M achieves superior long-term video generation performance over existing baselines, improving stability and visual fidelity, and establishing a new paradigm for learned memory management in long autoregressive video generation.

When does a language model acquire a capability? When its hidden states encode the relevant information, or when its output readout can express it? We investigate this gap between representational availability and readout expression by tracking the unembedding matrix through pretraining. Because unembedding rows are aligned by token identity across checkpoints, they provide one trajectory per vocabulary item through the learned output interface. To measure these trajectories, we introduce parameter-trajectory crosscoding, a checkpoint-spanning sparse dictionary with shared feature identities and checkpoint-specific decoders. Applied across Pythia scales and in OLMo-2-7B, this method reveals an early, heterogeneous reorganization of the readout. This reorganization does more than alter structure. Our independent WordNet probes show that token families become increasingly separable over the same developmental window. We then causally test whether the readout component drives this development. By swapping the unembedding matrices across checkpoints, and controlled contrastive tasks, we show that task-relevant distinctions can be present in hidden states before the contemporaneous readout expresses them in logits. By ablating individual crosscoder features, we trace these delayed readout effects directly to compact, learned directions. More broadly, developmental analyses should distinguish representational availability from readout expression. Together, our findings demonstrate that the learned output interface helps determine when latent structure becomes visible in token logits, highlighting the need for developmental analyses to explicitly decouple hidden-state representation from readout expression.

We study preference learning and coordination through recommendations in multi-agent game settings, where a moderator repeatedly interacts with agents whose utility functions are unknown. In each round, the moderator issues action recommendations and observes whether agents follow or deviate. We consider agents who best respond to the moderator's recommendations and study what this feedback reveals about their utilities. We characterize the class of games that are indistinguishable under this feedback model. Moreover, we introduce a notion of moderator regret based on agents' incentives to deviate from the recommendations and design an online algorithm with low regret under the best-response model, with guarantees that scale linearly in the game dimension and logarithmically in time. Our results lay a theoretical foundation for AI recommendation systems in strategic multi-agent environments, where recommendation compliance is shaped by strategic interaction.

Training tool-using agents for planning is challenging when high-quality demonstrations exist only as final outcomes without the intermediate tool traces that produced them. A common response is to impute traces and apply imitation learning, but this approach is indirect and structurally limited to matching the reference rather than improving upon it. We propose anchored comparative feedback: rather than imitating a reference plan, we train agents to surpass it under a comparative LLM judge while satisfying hard constraints. The anchor provides a stable optimization target, and comparing two identically rendered plans neutralizes surface-level judge biases. On TripTailor, a tool-augmented travel planning benchmark, anchored GRPO achieves 80.1% final success rate compared to 43.7% for imitation learning on imputed traces, with gains on both constraint satisfaction and LLM-judged preference that persist across judge families, rubrics, and a held-out reward model. Gains generalize beyond TripTailor, with a fivefold improvement on the held-out TravelPlanner benchmark. These results suggest that anchored comparative feedback offers an effective approach for learning transferable tool-using planners from outcome-only supervision.

Lipreading remains challenging due to the inherent ambiguity of visual speech cues and their sensitivity to head pose and video resolution. In this work, we propose FLARE, a framework for robust visual speech representation learning via cross-modal distillation and joint face-lip modeling. FLARE leverages semantically rich Whisper representations as the acoustic teacher signal, converting them into discrete distillation targets through a random-projection quantizer and pretraining the model with a masked prediction objective. To capture complementary visual information, FLARE employs a dual-stream architecture that jointly processes lip region-of-interest (ROI) and full-face inputs, where the lip stream captures fine-grained articulatory motion and the face stream encodes broader facial dynamics. During pretraining, auxiliary audio input is incorporated to facilitate cross-modal alignment, and modality dropout is applied to encourage robust visual-only representations. For downstream lipreading, FLARE serves as a visual encoder coupled with either a lightweight Transformer decoder or an LLM-based decoder. Evaluated on LRS2, LRS3, and the out-of-domain WildVSR benchmark, FLARE achieves state-of-the-art word error rates of 11.7%, 13.2%, and 34.3%, respectively. A progressive design study confirms the contribution of each major component. Further analyses demonstrate that joint face-lip modeling improves robustness under degraded video resolution and large head pose variation, while yielding more phoneme-discriminative visual representations. Code and pretrained models are available at https://anonymous.4open.science/r/FLARE-522C.

Machine unlearning aims to remove targeted knowledge from a trained model while preserving its general capabilities. Doing so effectively for auto-regressive language models requires identifying which tokens in a forget sample are truly relevant to forgetting, as uniformly applying unlearning across all tokens can degrade utility. Existing approaches, however, either ignore this heterogeneity or rely on auxiliary models, hand-crafted heuristics, or external annotations to approximate token-level relevance. We instead characterize token-level forget relevance through its interaction with the retain objective: tokens are forget-relevant to the extent that minimizing the forget loss on them does not conflict with the retain objective. Based on this insight, we introduce Alternating Token-Weighted Unlearning (ATWU), a framework that jointly learns token relevance and model parameters during the unlearning process. ATWU uses a lightweight linear scorer over model hidden states to predict forget-relevance scores with negligible computational overhead and no external supervision. Experiments on TOFU and RWKU show that ATWU achieves state-of-the-art forget--retain trade-offs, outperforming sample-level methods, probability-based token-weighting heuristics, and auxiliary-model-based approaches. Moreover, the learned scorer aligns with ground-truth forget-relevant spans substantially better than existing methods, suggesting that ATWU learns semantically meaningful token-level forgetting signals. Overall, ATWU shows that token-level forget-relevance can be effectively inferred from model representations during unlearning, enabling efficient and selective forgetting.

Behavior cloning provides strong imitation learning guarantees when training and test environments share the same dynamics. However, in many deployment settings the test environment's transitions differ from training, and classical offline IL offers no recourse: the learner must commit to an action at every state, even when its demonstrations are uninformative and could lead to arbitrary degradation of performance. This motivates the study of _selective_ imitation, where the learner may choose to _stop_ when it cannot act reliably. We introduce a model for selective imitation under arbitrary dynamics shift: given labeled expert demonstrations from a training environment and unlabeled state trajectories from the same expert in a test environment, the learner outputs a _selective policy_ that is _complete_ (rarely stops in training) and _sound_ (incurs low regret before stopping in test). Our algorithm, $\operatorname{SeqRejectron}$, constructs a stopping rule using a small set of _validator policies_ whose size is independent of the horizon or policy class. For deterministic policies, this yields horizon-free $\tilde{O}(\log|\Pi|/\epsilon^2)$ sample complexity, assuming sparse costs. For stochastic policies, we obtain analogous horizon-free guarantees using a cumulative Hellinger stopping time. We extend the framework to misspecified experts and different expert policies across train and test and obtain results that gracefully degrade with the amount of misspecification.

Large language models (LLMs) encode rich semantic knowledge that can be useful for supervised learning, but their outputs are unreliable as statistical priors: they may be noisy, misspecified, or hallucinated. Existing LLM-informed learning methods either trust such signals directly, leaving predictions vulnerable to unreliable LLM guidance, or restrict semantic integration to a single model class. We introduce Statsformer, a validated framework for learning when to trust LLM-derived semantic priors in supervised statistical learning. Statsformer maps LLM-derived feature scores into a family of learner-specific prior-injection mechanisms across a heterogeneous library of linear and nonlinear predictors. It then uses out-of-fold validation to adaptively calibrate the influence of each prior-informed learner, allowing useful semantic information to improve prediction while attenuating weak, misspecified, or adversarial priors. This yields a guardrailed statistical learning system with an oracle-style guarantee: up to statistical error, the final predictor performs no worse than the best convex combination of its in-library candidates, including prior-free learners. Across diverse prediction tasks, informative LLM priors improve performance, while unreliable priors are automatically downweighted. These results position Statsformer as a reliability-oriented approach to LLM-informed statistical learning: rather than trusting LLM knowledge directly, it validates semantic priors against data before allowing them to influence the final predictor.

Preference optimization aligns large language models from pairwise preferences by increasing the margin between preferred and dispreferred responses. However, margin-based losses alone do not test whether each pair's preference signal is locally stable enough to trust with full update strength. We propose Geometric Anchor Preference Optimization (GAPO), a geometry-aware objective that introduces a batch-conditioned stress test for preference learning. For each mini-batch, GAPO constructs a pessimistic anchor by perturbing the current policy in the first-order direction that decreases the batch-average preference margin. The resulting Anchor Gap measures how much each pair's margin degrades under this shared pessimistic probe and converts this degradation into an instance-wise update weight. Pairs with larger Anchor Gap receive smaller update weights, while non-brittle pairs retain their preference-gradient direction. Under smoothness assumptions, we characterize the Anchor Gap as a batch-directional proxy for local margin degradation. Across multiple open-weight model families, GAPO matches or improves strong preference-optimization baselines on instruction-following and reasoning benchmarks. It also improves robustness under random and structured preference noise without explicitly modeling label corruption. Mechanistic diagnostics show that the pairs receiving the smallest GAPO weights are statistically enriched in corrupted supervision, suggesting that GAPO improves robustness by reducing the cumulative influence of brittle preference signals.

Representation learning for single-cell RNA-seq has been evaluated by the identities representations capture — cell types, lineages, perturbations — rather than how faithfully they model the distribution of cell states on the gene expression manifold. As a result, density estimation, rare-state detection, perturbation scoring, and gene-level attribution are bolted onto frozen embeddings as separate models, disconnected from the geometry of the data. We argue that the density model should be the primary object and that representation quality follows from it. Building on LeJEPA, whose SIGReg objective enforces a provably optimal isotropic Gaussian latent, we introduce LeCellModel, which exploits this Gaussian latent to recover data density on the gene expression manifold via the encoder's Jacobian log-volume (JEPA-SCORE). Across several datasets, LeCellModel outperforms established methods on scIB benchmark for representation quality. Furthermore, we show that LeCellModel enables estimation of rare expression states using JEPA-SCORE as a typicality measure – applying it to a controlled cytokine panel, where it identifies stimulated cells across minority fractions and tracks pathway-response magnitude. We introduce gene-level interpretability of density estimates — to our knowledge, the first method for analyzing JEPA-SCORE in feature space. Leveraging this framework, LeCellModel recapitulates the known association of a profibrotic macrophage program with idiopathic pulmonary fibrosis, and discovers a novel CD177$^{+}$ activated regulatory T cell subpopulation previously described only in tumor contexts. Both findings emerge directly from the geometry of the learned embedding, without relying on supervised annotations.

We propose LEGO, a budget-aware Dense-to-MoE methodology that converts pretrained vision-language models into efficient FFN-MoE variants under a target architectural budget $((P_{total},P_{active}))$. Our study is motivated by a counter-intuitive recovery pattern: fixed activation ratios that work well for some dense backbones can degrade sharply on others under the same recovery recipe, indicating that raw sparsity alone is an unreliable design rule. Through controlled recovery experiments across VLM families and scales, we identify the activated-FFN-to-hidden ratio $(D_I/D_H)$ as a backbone-comparable structural factor and summarize empirical sizing rules for avoiding severe bottlenecks while limiting diminishing returns. LEGO operationalizes these rules through a discrete budget-aware search over Split, Upcycle, and Hybrid constructions, and introduces a moment-matching scaling factor to reduce the initialization scale shift caused by sparse FFN aggregation. Finally, we use a two-stage multimodal recovery recipe that first stabilizes the sparse LLM with the vision tower frozen, then performs joint fine-tuning to recover the full VLM. Under matched data and training protocols, LEGO improves Dense-to-MoE recovery quality and turns costly blind architecture search into a small set of rule-guided candidates. Codes can be found in supplementary materials.

Robust and accurate neural decoders are integral to the development of neurotechnological applications such as brain-computer interfaces and closed-loop experiments. Recent work has shown that tokenizing neural data at the resolution of individual spikes facilitates multi-session pretraining and delivers state-of-the-art decoding performance. However, these spike-based models are currently restricted to supervised learning (SL) regimes, limiting both pretraining and finetuning to datasets with paired behavioural labels. To address this, we introduce MOJO ($\textbf{M}$asked aut$\textbf{O}$encoder-based $\textbf{JO}$int training), a training framework designed for spike-tokenizing models that jointly leverages self-supervised learning (SSL) via masked autoencoding and SL objectives for better decoding performance and interpretability. We evaluate MOJO on three intracortical spiking datasets—monkey motor cortex across various reaching tasks, multi-regional mouse recordings during vision and cognitive decision making tasks—demonstrating superior performance to purely SL-trained models. This improvement is especially notable when training with limited labelled data, specifically in the few-shot finetuning regime where only a small amount of labelled data is available to adapt a model to a new recording session. The incorporation of SSL also yields more interpretable neuronal representations, improving performance on analyses such as brain region classification and spike-statistics prediction despite the lack of explicit optimization for these tasks. We then show that MOJO generalizes beyond spiking data using human electrocorticography (ECoG) during speech articulation. We show that MOJO continues to outperform purely SL-trained models and achieves performance comparable to neuro-foundation models (NFMs) built specifically for continuous signals. Overall, augmenting spike-tokenizing models with SSL improves their performance in label-impoverished settings and enables the use of unlabelled data across various tasks and species, and while generalizing to other neural modalities. These results suggest a path towards more flexible and scalable data usage when training NFMs.

Constructive neural routing solvers usually score the next action by matching a decoder context to candidate embeddings, hiding deterministic one-step consequences such as travel, waiting, slack, and capacity changes. We propose LINC (Local Inference via Normed Comparison), a decoder-side candidate decision architecture that computes these consequences explicitly. LINC uses them according to their decision role: centered relative consequences are compared by a shared linear local scorer, while feasible-set summaries modulate the decoder context. This preserves standard global matching and relieves the hidden state from rediscovering transition arithmetic. The Capacitated Vehicle Routing Problem with Time Windows (CVRPTW) serves as the main constrained-routing stress test; the same interface extends to the Capacitated Vehicle Routing Problem (CVRP) and Traveling Salesman Problem (TSP). In particular, for CVRPTW, LINC reduces PolyNet's Solomon/Homberger gaps from 13.83\%/38.15\% to 7.26\%/14.71\%; for TSP and CVRP, it also improves external-benchmark gaps.

Ensemble sampling is a practically attractive approach to randomized exploration, but existing theoretical guarantees for linear bandits require ensembles much larger than what its practical motivation would suggest. In particular, the sharpest existing analysis achieves the ensemble complexity of $\Theta(d\log T)$, leaving a $\log T$ gap from the intrinsic $\Omega(d)$ ensemble-size barrier. We aim to narrow this gap by proposing an ensemble sampling algorithm that refreshes the ensemble only when the regularized Gram matrix changes substantially. This mechanism localizes the perturbation analysis to epochs with controlled Gram-matrix drift and reduces the sufficient ensemble size to $\Theta(d\log d+d\log\log T)$, while preserving the state-of-the-art $\tilde{\mathcal O}(d^{3/2}\sqrt T)$ regret for ensemble sampling with arbitrary bounded arm sets. We further show that, when the arm set is finite of cardinality $K$, the proposed algorithm achieves the sharper regret bound $\tilde{\mathcal{O}}(d\sqrt{T\log K})$. To the best of our knowledge, this is the first ensemble-sampling guarantee that simultaneously recovers both canonical regret scalings known for randomized linear bandit algorithms: the $\tilde{\mathcal{O}}(d^{3/2}\sqrt{T})$ rate for arbitrary bounded arm sets and the $\tilde{\mathcal{O}}(d\sqrt{T\log K})$ rate for finite arm sets. The algorithm also admits an anytime implementation without resetting past data, and experiments show that it remains competitive with baselines while using substantially smaller ensembles.