Risk-Calibrated Proposal Transport for Finite-Particle Diffusion Steering

Inference-time steering combines pretrained diffusion experts or rewards without retraining. Linear combinations of expert fields yield a cheap unweighted Anchor but generally transport a biased law; Feynman-Kac correction recovers the target asymptotically through importance-weighted Sequential Monte Carlo, whose finite-particle behavior depends on the proposal. Variance-controlling guidance (VCG) addresses this by fitting a linear controller to minimize empirical Feynman-Kac log-weight-rate variance, aiming to reduce both unweighted transport error and correction variance. Despite its ideal population (i.e., infinite-particle limit) solution being non-worsening, VCG minimizes variance on a finite ensemble in practice, and this gap can be catastrophic: empirical VCG may nearly eliminate its fitting residual yet amplify residual on new states by orders of magnitude, collapsing a weighted particle system to one or a few ancestral lineages or severely biasing unweighted generation. To address this, we first show that the centered Feynman-Kac rate is the normalized transport residual and that out-of-fit benefit is exactly population headroom minus coefficient-estimation penalty. We introduce Risk-Calibrated Proposal Transport (RCPT), a nearly cost-free, plug-and-play method that estimates this tradeoff with deletion leave-one-out residuals and retains only the supported portion of the VCG update. It reuses the same particles, fields, and expert evaluations, adding no model calls and only small linear algebra. On Checker, VCG reaches an independent residual $94.0\times$ that of Anchor despite a near-zero fitting residual. RCPT cuts it by $99.1\\%$ and finishes below Anchor. Across a 76-pocket scaffold benchmark and multi-reward molecular SMC, RCPT also lowers transport risk, restores downstream molecular quality, and repairs particle genealogy. As fitting information grows, RCPT recovers VCG.