Recent progress in large reasoning models (LRMs), like OpenAI o1 and Deepseek R1, has been pivotal for tackling complex applications, from mathematical and code reasoning to advanced symbolic and agentic planning. Their success often relies on test-time scaling, which involves increasing the generation length or depth. However, these approaches incur significant efficiency bottlenecks during training and inference. To overcome these limitations, further advancements are needed in data, algorithms, and systems applicable across various domains, as exemplified by work such as s1, Z1, and verl. The proposed workshop will bring together researchers and practitioners to rethink efficient reasoning under tight compute, memory, latency, throughput, and cost budgets, with the goal of translating theoretical breakthroughs into practical, deployable solutions.

Much of traditional decision-making science is grounded in the mathematical formulations and analyses of structured systems to recommend decisions that are optimized, robust, and uncertainty-aware. This scientific approach, rooted in the field of Operations Research (OR), has evolved through decades of advancements in stochastic modeling, computational simulation and optimization, and exhibits key strengths in methodological rigor and uncertainty encoding. On the other hand, recent advances in the AI/ML space have eschewed this model-based paradigm and increasingly embraced, to great success, model-free algorithmic design frameworks. This workshop, which is the first NeurIPS workshop explicitly themed and structured on ML-OR synergization, aspires to present recent developments, challenges and emerging research to accelerate ML-OR synthesis. By integrating ML into established OR methodologies, we have the opportunities to produce more data-centric and adaptive solutions for complex decision-making tasks that could propel, in a much faster-paced manner, the frontier of "optimality" across many relevant applications. Concomitantly, the goal is also to explore how model-based principled OR approaches can help alleviate issues revolving around "black box" systems, and provide paths to enhance interpretability, trust, and performance analysis.

This competition focuses on advancing fairness-aware detection of AI-generated (deepfake) faces and promoting new methodological innovations, addressing a critical gap where fairness methods developed in machine learning have been largely overlooked in deepfake detection. In the competition, participants will work with two large-scale datasets provided by the organizers: AI-Face (CVPR 2025), a million-scale, demographically annotated dataset for training and validation, and PDID (AAAI 2024), a newly curated dataset comprising real-world deepfake incidents, reserved for testing. Participants are tasked with developing models that achieve strong utility performance (e.g., AUC) while ensuring fairness generalization under real-world deployment conditions. The baseline method, PG-FDD (published at CVPR 2024 from the organizer’s group), which demonstrates state-of-the-art performance in fairness generalization for AI face detection, will be provided to support participation.The competition’s potential impact includes fostering the development of robust, fair, and generalizable deepfake detectors, raising awareness of fairness challenges in combating AI-generated fakes, and promoting responsible AI and machine learning deployment in societal applications such as media forensics and digital identity verification. Our competition is fortunately sponsored by Deep Media AI and Originality.AI companies. The challenge link is https://sites.google.com/view/aifacedetection/home.

This workshop aims to explore how real-world advances in video generation increasingly rely on forwardlooking evaluation frameworks and to collaboratively shape the next generation of high-quality video synthesis. Through a combination of invited talks, academic presentations, and expert discussions featuring leading voices from both academia and industry, the workshop bridges academic foundations and industrial insights across the modeling, evaluation, and deployment of video generation. We welcome contributions from computer vision, generative modeling, video-language learning, evaluation methodology, and human-centered AI to shape the next generation of high-quality video synthesis collaboratively.

The field of wireless communications and networking is undergoing a paradigm shift, driven by the growing potential of Artificial Intelligence (AI) and Machine Learning (ML) to redefine traditional system design principles. This workshop aims to catalyze interest and foster collaboration between the AI/ML and wireless communications communities. The timing of this workshop is especially significant, as the next-generation (NextG) wireless standardization efforts (such as 6G and WiFi 9) are just getting started, with AI-native technologies expected to play a central role across all aspects of the wireless ecosystem – from radio access to network management and edge intelligence. NextG represents a foundational shift in global infrastructure, enabling ultra-fast, low-latency, and intelligent connectivity that will power future applications in AI, robotics, immersive environments, and autonomous systems. These technologies offer unprecedented opportunities to both drive and benefit many applications, from healthcare and transportation to industrial automation and environmental monitoring. The economic and societal implications are vast: NextG networks will underlie trillions in global GDP impact, bridge digital divides, and shape how billions of people interact with technology and each other in the decades to come.

Despite the clear promise, a significant disconnect exists between the AI/ML and wireless research communities. AI/ML experts often lack an understanding of the unique physical, algorithmic, and architectural constraints inherent in wireless systems, while wireless researchers tend to adopt generic, off-the-shelf AI/ML models that are not optimized for the intricacies of wireless environments. Wireless environments are inherently dynamic, high-dimensional, and partially observable. These unique characteristics make them ideal testbeds for developing robust learning algorithms, particularly in areas like online learning, reinforcement learning, and in-context learning. At the same time, AI/ML techniques are becoming essential for managing the growing complexity of modern wireless networks, including resource allocation, interference mitigation, and cross-layer optimization. Bridging the gap between the two communities is not only necessary for meaningful technological advances but also critical for realizing the full societal impact of intelligent wireless systems.

This workshop aims to bring together researchers and practitioners at the intersection of artificial intelligence (AI), machine learning (ML), and wireless to address the unique challenges and opportunities posed by Next-Generation (NextG) wireless systems. As the 6G era begins to take shape, AI-native designs have emerged as a cornerstone of wireless innovation, with the potential to transform the performance, efficiency, and adaptability of communication systems. The integration of AI/ML is poised to influence every layer of the network stack, from physical-layer signal processing to network control and resource management.

The study of collective action has a long history in economics and sociology as a way for groups of people to impact markets and the political arena. Algorithmic collective action focuses on the study of such coordinated efforts in algorithmically-mediated sociotechnical systems. How can participants of AI systems coordinate towards a common good? We offer a platform to discuss new ideas and help define the foundational research directions for this emerging topic through interdisciplinary discussions between core ML researchers, scholars from the social sciences, community stakeholders and advocates.

The past few years have seen an unprecedented surge in both the availability of bioacoustic data and the sophistication of AI/machine learning models. This convergence presents a unique window of opportunity to revolutionize our understanding of animal communication and biodiversity. However, achieving this requires a conscious effort to integrate the disciplines of AI/Machine Learning and Ethology.

This workshop will explore the intersection of artificial intelligence (AI) and bioacoustics, aiming to address challenges in processing complex bioacoustic data and interpreting animal signals in order to advance our understanding of non-human animal communication. Join us for a poster session, keynote talks and a panel discussion as we explore key opportunities to use AI to decipher animal languages and thus deepen our understanding of the natural world.

Athletic performance represents the pinnacle of human decision-making. It demands rapid choices, precise motor control, agility, and coordinated physical execution. Such a combination of capabilities remains elusive in current artificial intelligence and robotic systems.Building on the momentum of the MyoChallenge at NeurIPS 2022, 2023, and 2024, the 4th edition of our MyoChallenge series -- Towards Human Athletic Intelligence-- moves toward capturing the full expressivity and agility of human athletic performance. Participants will develop behaviors for physiologically realistic musculoskeletal models performing fast-paced, and high-skill athletic tasks.The challenge will feature two tracks. First, a Soccer shootout. A full-body musculoskeletal model must dynamically approach and shoot a ball past a moving goalkeeper. Success requires balance, foot targeting, force generation, and rapid whole-body coordination. Second, a Table Tennis competition. A musculoskeletal model of the upper body (arm and trunk) must track, strike, and return balls in a fast-paced table tennis rally against an AI opponent. These challenges go far beyond static or repetitive motions. They demand generalization via reactive and adaptive embodied behavior grounded in the physics of muscle, tendon, and joint dynamics, with real-time perception-action loops capable of agile motor control. The challenge will be staged in the commonly used MyoSuite framework, which offers physiologically accurate, state-of-the-art musculoskeletal models, an intuitive interface to scalable reinforcement learning and control libraries. The framework also enables easy onboarding via extensive tutorials and getting-started materials, and access to multiple baseline libraries needed for the challenge.The competition aims to engage diverse research communities: biomechanics, motor neuroscience, reinforcement learning, control theory, and more. As in previous years, it will prioritize scalability, reproducibility, and generalization, and be open-sourced following best engineering and academic practices to advance physiological control and bring us closer to replicating human athletic intelligence.

AI4Mat-NeurIPS-2025 explores applications of artificial intelligence (AI) to materials via: 1. AI-Guided Materials Design; 2. Automated Chemical Synthesis; 3. Automated Material Characterization. AI4MatNeurIPS-2025 emphasizes structured, expert-driven dialogue on making advanced machine learning more impactful for real-world materials discovery. To that end, AI4Mat-RLSF (Research Learning from Speaker Feedback) creates a new structured discussion format where spotlight presenters receive curated, in-depth feedback from invited discussants. Further, the AI4Mat Frontiers & Benchmarking session brings together a diverse and distinguished set of speakers to critically examine current benchmarks, present state-of-the-art methods, and identify emerging opportunities and current limitations in AI-driven materials design.

As the US FDA phases out animal testing requirements for drug discovery and development, AI tools will become widely adopted to simulate the effects of candidate drugs. We posit that building virtual cells and instruments with AI is poised to transform drug discovery and development by enabling large-scale simulation and interrogation of molecules, cells, and tissues. In our workshop, we will collaboratively define and promote this emerging scientific paradigm of AI to accelerate the drug discovery and development process in this new era.

Recent progress in reinforcement learning (RL) has powered breakthroughs in various real-world problems, gathering considerable attention and investment. However, it has also exposed a significant gap between theoretical and experimental developments.

RL theory has grown significantly in the past two decades. Research has characterized the inherent difficulty of various settings and has designed a wide variety of algorithms to reach optimal performances. Furthermore, a huge leap has been made in understanding how to handle large state spaces using function approximation techniques, identifying key structural properties that enable efficient learning.

Despite theoretical guarantees, applying RL algorithms to complex problems faces challenges. Theoretical algorithms often focus on simplified settings, making them hard to apply to real-world complexities. Furthermore, optimizing for worst-case scenarios, which include unlikely situations, can lead to algorithms that perform poorly on practical tasks. Yet, while specialized algorithms offer empirical success, they might not translate to other problems due to their specific design, and the reliance on heuristics and engineering fixes further widens the gap between theory and practice.

A prominent area that has seen a surge of interest in RL is generative language modeling. Pre-training these models can be viewed as a form of imitation learning, while post-training typically implements RL algorithms for purposes like instruction tuning with RL from human feedback or enhancing reasoning capabilities. While these successes make the practical utility of RL undeniable, the RL community finds itself at a crossroads. The algorithms employed are frequently variants of classical methods, and exploring beyond these presents a key challenge. Conversely, the success of these models prompts new questions for RL theory, suggesting that frameworks leveraging pre-trained models might offer a more effective paradigm than learning from scratch under traditional assumptions.

Following the success of the ICML 2024 edition, the Second Workshop on Aligning Reinforcement Learning Experimentalists and Theorists (ARLET) aims to bridge this discrepancy and promote collaboration. By bringing together experts from both sides, we want to facilitate meaningful discussions and draw a path for future RL research. Motivated by the take-home messages from the previous edition, we seek to encourage: (i) theorists to ask experimentalists for concrete problems to solve, and (ii) experimentalists to seek theoretical guidance on how to approach these problems.

Our brains and bodies speak a rich and complex biological language of neural and physiological signals, a language that AI models are increasingly capable of deciphering as large-scale datasets become available. Recent advances in brain interfacing and wearable technologies, including EEG, intracortical electrophysiology, EMG, MEG, and ECG, have enabled the broad collection of these signals across real-world contexts and diverse populations. This growing wealth of data is driving a shift toward foundation models: large-scale, pretrained AI systems designed to learn from biosignals and generalize across diverse downstream applications, from brain-computer interfacing to health monitoring. Realizing this potential, however, requires addressing the unique challenges that come with biosignal timeseries: they are noisy, heterogeneous, and collected under variable conditions across subjects, devices, and environments. To meet these challenges, this workshop brings together neuroscientists, biomedical engineers, wearable tech researchers, and machine learning experts advancing foundation model approaches. Through interdisciplinary dialogue, we aim to catalyze the next generation of AI models that can capture the complexity of the brain, body, and behavior at scale.

Deep learning for code has progressed from focused tasks—such as completion, repair, synthesis, and explanation to tackling complex, end-to-end software–engineering problems. A key recent breakthrough is the rise of coding agents. Unlike single-shot models, these systems plan, reason, explore, and invoke external tools to assist throughout the software-development lifecycle: adding features, refactoring, debugging, finding vulnerabilities, optimizing performance, summarizing code, and answering repository-level questions. Their growing versatility demands rigorous evaluation and a deeper understanding of their capabilities, limits, risks, and broader social impact.

Building on momentum from both academia and industry (e.g. Google, OpenAI, Anthropic, SWE-Agent, OpenHands), we propose the 4th Deep Learning for Code (DL4C) workshop with a dedicated focus on coding agents. This workshop will provide a timely forum where researchers and practitioners can design and stress-test robust coding agents, discover novel applications and emergent behaviors, establish principled benchmarks and evaluation methods, study human–agent collaboration at scale, and advance the responsible, safe deployment of autonomous coding tools.

World models infer and predict real-world dynamics by modeling the external environment, and have become a cornerstone of embodied artificial intelligence. They have powered recent progress in decision-making and planning for interacting agents. This workshop aims to bring together researchers working at the intersection of generative modeling, reinforcement learning, computer vision, and robotics to explore the next generation of embodied world models—models that enable agents to understand, predict, and interact with the world through learned models. By focusing on embodiment and decision-making, this workshop seeks to advance world models beyond passive prediction, toward active, goal-driven interaction with the physical and virtual world. By emphasizing embodiment and decision-making, we aim to move beyond passive sequence prediction toward goal-directed interaction with both physical and simulated worlds.

This workshop aims to foster cross-disciplinary collaboration at the intersection of generative AI and finance, a high-stakes domain where the integration of domain expertise is essential to the safe and effective deployment of machine learning technologies. Recent advances in generative models—ranging from large language models to diffusion and score-based generative architectures—have opened new frontiers for applications in finance, such as financial modeling, stress testing, scenario generation, automated financial services, and decision-making under uncertainty.

The workshop will highlight theoretical advances, practical implementations, new opportunities, and open challenges that arise when adapting generative AI to financial systems under unique constraints, such as data sparsity, regulatory requirements, and highly non-stationary and adversarial environments. By bringing together the computer science community, financial researchers, industry practitioners, and regulators, we aim to catalyze interdisciplinary dialogue and accelerate the responsible development of generative AI tailored to the needs of finance and risk management.

Generative AI (GenAI) has emerged as a transformative force in healthcare, yet public trust remains limited due to safety concerns and policy misalignment. Build- ing on last year’s successful GenAI4Health workshop, the field has rapidly evolved from exploratory research to real-world clinical deployments, accompanied by FDA regulatory involvement and expanding global governance frameworks. This second workshop convenes machine learning researchers, healthcare professionals, and policy experts to address the critical intersection of GenAI innovation and regula- tory compliance in health applications. We will examine trustworthiness challenges in large language models and multimodal foundation models, explore mitigation strategies, and foster dialogue between technical and policy communities. Our goal is to advance safe, effective, and ethically-compliant GenAI integration in healthcare systems, improving patient outcomes and clinical research capabilities

Recent advancements in generative AI (GenAI) have empowered machines to create high-quality content across diverse modalities - text, image, audio, and video - with impressive fluency and creativity. From GPT-4o and Stable Diffusion to Sora and MMAudio, the explosion of X-to-X generation (e.g., text-to-image, video-to-audio) is unlocking new frontiers in science, education, entertainment, and art.

While GenAI has shown significant potential in creative applications (e.g., music, films, arts), these breakthroughs also raise pressing concerns related to safety, copyright, and ethical use. Generative models can be exploited to spread misinformation, violate intellectual property rights, or diminish human agency in creative processes. As such, there is an increasing need to balance innovation with protection, ensuring that AI-powered creative tools are used responsibly and ethically.

This workshop, GenProCC: Generative and Protective AI for Content Creation, brings together researchers, creators, and practitioners at the intersection of content generation and IP protection. By uniting the generative AI and creator communities, the GenProCC workshop aims to explore the latest advances, challenges, and opportunities in the rapidly evolving field.

Imageomics is an emerging interdisciplinary field at the crossroads of machine learning (ML), computer vision (CV), and biological sciences. It leverages visual data—from microscopic images of single-cell species to videos of megafauna—to extract and analyze biological information, specifically traits. By grounding ML models in existing scientific knowledge, Imageomics aims to make traits computable from images, facilitating insights into the evolution and function of living organisms. Imageomics poses research problems that resonate with the broad machine-learning community: multimodal representation learning, object detection and tracking, few-shot learning, imbalanced-class learning, video understanding, 3D modeling, hierarchical learning, etc. When people leverage ML tools to solve biological questions, the foundational bridges between ML and biological sciences also provide opportunities to address key challenges in ML, creating a virtuous cycle between the two fields.

Large Language Models (LLMs) have emerged as transformative tools across research and industry, revolutionizing how we interact with information. However, their immense capabilities bring critical security challenges—the same features that drive innovation can be exploited for malicious purposes through unauthorized distillation, fine-tuning, compression, or editing. These vulnerabilities pose severe threats, including intellectual property theft, the generation of sophisticated disinformation, the bypass of safety alignments, and the erosion of user trust in AI systems.

This workshop aims to bring together researchers and practitioners from academia and industry who are advancing the frontiers of LLM security and protection. We seek to confront the unauthorized use of LLMs head-on by exploring novel and robust mechanisms designed to make these models inherently resistant to exploitation while maintaining their beneficial capabilities. The workshop also hosts the 2025 TrustAI Rising Star Award.

Topics of interest include, but are not limited to:
1. Un-Distillable LLMs: Preventing unauthorized model replication and intellectual property theft
2. Un-Finetunable LLMs: Resisting malicious parameter updates and behavior alterations
3. Un-Compressible LLMs: Maintaining model integrity against unauthorized compression
4. Un-Editable LLMs: Safeguarding against knowledge tampering and misinformation injection
5. Un-Usable LLMs: Ensuring traceability and preventing misuse through watermarking and verification

The Machine Learning and the Physical Sciences (ML4PS) workshop at NeurIPS is a unique gathering space for the growing community spearheading cross-cutting research topics at the intersection of machine learning (ML) and the physical sciences (PS). This includes the applications of ML to problems in the physical sciences (ML for PS) as well as developments in ML motivated by physical insights (PS for ML). The physical sciences are defined inclusively, including but not limited to physics, astronomy, cosmology, chemistry, biophysics, materials science, and Earth science. Join us to discuss the latest research at the convergence of these fields!

The 9th Machine Learning for Systems (ML for Systems) workshop brings together researchers and practitioners applying machine learning to core computer systems challenges. This year, we focus on three themes: (1) using LLMs and agentic workflows for systems tasks such as program synthesis and adaptive optimization; (2) applying ML to manage the complexity of large-scale training and serving of multimodal and reasoning models; and (3) leveraging ML for sustainable computing, including energy-, power-, and carbon-aware optimization. The workshop will feature invited talks, contributed presentations, and discussions aimed at advancing the frontier of ML for Systems research.

The field of AI is entering a new era of interaction, profoundly shaped by the capabilities of Large Language Models (LLMs). While multi-turn interaction has been a long-standing pursuit in AI—from dialogue systems to multi-agent coordination—the advent of LLMs has radically transformed this landscape. These models now engage in complex, long-horizon interactions, process diverse data, and make crucial decisions in dynamic, human-centric scenarios.

This leap forward, however, brings forth critical new research questions and challenges that demand immediate attention:

Multi-Turn RL Learning for Agentic Tasks Learning from complex, interactive environments like GUI agents and tool-use scenarios, given the challenges of sparse rewards.
Maintaining Alignment Understanding human values over extended, multi-turn interactions, preventing "loss of alignment" seen in current models.
Human-AI Interaction Over time, ensuring models adapt to user goals without compromising safety or fairness.
Long-horizon Evaluation For LLMs' long-term capabilities, consistency, and strategic abilities in complex, multi-turn tasks.
The Workshop on Multi-Turn Interactions in LLMs is designed to be the central forum for addressing these pivotal questions. We invite researchers to contribute to defining the next generation of interactive AI, tackling these core challenges, and charting the course for future advancements in AI reasoning and planning. This workshop will concentrate on key areas where the extended use of LLMs presents both new challenges and opportunities, serving as a platform to discuss and refine methods for future improvements and evaluation for practical LLM use cases.

Optimization lies at the heart of many machine learning algorithms and enjoys great interest in our community. Indeed, this intimate relation of optimization with ML is the key motivation for the OPT series of workshops. We aim to foster discussion, discovery, and dissemination of state-of-the-art research in optimization relevant to ML.

The focus of OPT 2025 is on "Statistics Meets Optimization". Since its inception, stochastic optimization has been grounded in statistical principles. Today, many of the most pressing challenges in machine learning—such as generalization bounds, the training dynamics of overparameterized models, and the development of generative models—are directly inspired by statistical thinking. At the same time, the scale and complexity of modern datasets, along with the increasingly rich model classes used to represent them, pose new questions about how optimization algorithms interact with these structures—both computationally and statistically. For example, what role do data symmetries play in shaping optimization trajectories? How do statistical properties of the data affect the adaptivity and efficiency of learning algorithms? And how can optimization approaches be designed to scale with data while still preserving desirable statistical behavior? OPT 2025 will explore these questions with the goal of building bridges between the statistics and optimization communities, and highlighting their shared impact on the theory and practice of machine learning.

We are looking forward to seeing you all at OPT 2025, which will take place at the San Diego Convention Center!

Distributions shift, chatbots get jail‑broken, users game algorithms — how do we build reliable machine learning when data are missing, corrupted, or strategically manipulated?

This workshop bridges theory and practice to tackle these challenges, bringing together researchers working on distribution shift, adversarial robustness, and strategic behaviour to chart principled yet deployable solutions for Reliable ML from Unreliable Data.

When, how and why do different neural models learn the same representations?

New findings in neuroscience and artificial intelligence reveal a shared pattern: whether in biological brains or artificial models, different learning systems tend to create similar representations when subject to similar stimuli.

The emergence of these similar representations is igniting a growing interest in the fields of neuroscience and artificial intelligence, with both fields offering promising directions for their theoretical understanding. These include analyzing the learning dynamics in neuroscience and studying the problem of identifiability in the functional and parameter space in artificial intelligence.

While the theoretical aspects already demand investigation, the practical applications are equally compelling: aligning representations allows for model merging, stitching and reuse, while also playing a crucial role in multi-modal scenarios. Furthermore, studying the features that are universally highlighted by different learning processes brings us closer to pinpoint the invariances that naturally emerge from learning models, possibly suggesting ways to enforce them.

The objective of the workshop is to discuss theoretical findings, empirical evidence and practical applications of this phenomenon, benefiting from the cross-pollination of different fields (ML, Neuroscience, Cognitive Science) to foster the exchange of ideas and encourage collaborations.

In conclusion, our primary focus is to delve into the underlying reasons, mechanisms, and extent of similarity in internal representations across distinct neural models, with the ultimate goal of unifying them into a single cohesive whole.

Current electroencephalogram (EEG) decoding models are typically trained on specific subjects and specific tasks. Here, we introduce a large-scale, code-submission-based competition to subsume this approach through two challenges. First, the transfer challenge consists of building a model that can zero-shot decode new tasks and new subjects from their EEG. Second, the psychopathology factor prediction challenge consists of predicting measures of mental health from EEG data. For this, we use an unprecedented, multi-terabyte dataset of high-density EEG signals (128 channels) recorded from over 3,000 subjects engaged in multiple active and passive tasks. We provide several tunable neural network baselines for each of these two challenges, including a simple network and demographic-based regression models. Developing models that generalize across tasks and individuals will pave the way for EEG architectures capable of adapting to diverse tasks and individuals. Similarly, predicting mental health dimensions from EEG will be essential to systematically identify objective biomarkers for clinical diagnosis and personalized treatment. Ultimately, the advances spurred by this challenge are poised to shape the future of neurotechnology and computational psychiatry, catalyzing breakthroughs in both fundamental neuroscience and applied clinical research.

We propose a new data-centric competition that aims to advance the visual reasoning capabilities of vision-language models (VLMs) through instruction-tuning dataset curation. Participants are provided with a pool of 1 million image-text pairs and tasked with generating a small (1K) or large (10K) instruction-tuning dataset using any method of their choice. Submissions will be evaluated by fine-tuning a fixed VLM (Molmo) on the curated data and measuring performance on VMCBench, a newly released benchmark composed of multiple-choice visual reasoning questions spanning six diverse datasets.The competition provides all necessary resources, including the image-text pool, fine-tuning scripts, evaluation code, and baselines generated using GPT-4o and Claude, as well as 400 USD GPU compute from Lambda Labs. The evaluation metric is accuracy, and all training and evaluation will be reproduced by organizers on standardized infrastructure. This challenge reframes data curation as the primary variable for scientific investigation, with implications for adapting foundation models to real-world domains such as education, biomedicine, and scientific reasoning.We aim to foster broad participation across academia and industry, democratizing model adaptation by focusing on data quality rather than computational scale.

Precision therapeutics require models that can reason over complex relationships between patients, diseases, and drugs. Large language models and large reasoning models, especially when combined with external tool use and multi-agent coordination, have demonstrated the potential to perform structured, multi-step reasoning in clinical settings. However, existing benchmarks (mostly QA benchmarks) do not evaluate these capabilities in the context of real-world therapeutic decision-making. We present CURE-Bench, a competition and benchmark for evaluating AI models in drug decision-making and treatment planning. CURE-Bench includes clinically grounded tasks such as recommending treatments, assessing drug safety and efficacy, designing treatment plans, and identifying repurposing opportunities for diseases with limited therapeutic options. The competition has two tracks: one for models reasoning using internal knowledge, and another one for agentic reasoning that integrates external tools and real-time information. Evaluation data are generated using a validated multi-agent pipeline that produces realistic questions, reasoning traces, and tool-based solutions. Participants will have access to baselines spanning both open-weight and API-based models, along with standardized metrics for correctness, factuality, interpretability, and robustness. Human expert evaluation provides an additional layer of validation. CURE-Bench provides a rigorous, reproducible competition framework for assessing the performance, robustness, and interpretability of reasoning models in high-stakes clinical applications. It will accelerate the development of therapeutic AI and foster collaboration between AI and therapeutics communities.

Machine learning (ML) holds immense potential for discovering sustainable polymer materials, yet progress is hindered by the lack of high-quality open data. We provide an open-sourced dataset that is ten times larger than existing ones, along with competitive ML baselines and evaluation pipelines. This challenge targets multi-task polymer property prediction, which is crucial for virtual screening of polymers.Participants are asked to develop accurate prediction models, with a focus on material properties. A variety of ML techniques such as data augmentation and imbalanced learning, sophisticated learning paradigms like transfer learning and self-supervised learning, and novel model architectures with a good inductive bias on polymers can be leveraged. The competition results will directly accelerate the discovery of novel polymers for sustainable and energy-saving materials.