Risk-Calibrated Proposal Transport for Finite-Particle Diffusion Steering
Abstract
Inference-time steering combines pretrained diffusion experts or rewards without retraining by changing the dynamics that transport noise to data. Feynman-Kac correction compensates for proposal mismatch through importance-weighted sequential Monte Carlo (SMC), whose finite-particle behavior depends on the proposal. Variance-controlling guidance (VCG) improves that proposal by fitting a linear drift correction to minimize empirical log-weight-rate variance. Although its population optimum cannot worsen residual variance, finite-particle VCG can nearly eliminate its fitting residual while increasing residual risk on new states by orders of magnitude. The resulting update can degrade unweighted generation or accelerate particle collapse. We show that the centered Feynman-Kac rate is the normalized transport residual and that expected out-of-fit benefit is exactly population headroom minus coefficient-estimation penalty. Under regularity assumptions, a Wasserstein analysis bounds the unweighted proposal's terminal error using this residual. These results motivate Risk-Calibrated Proposal Transport (RCPT), which uses deletion leave-one-out residuals to calibrate the retained fraction of the VCG update, adding no model calls and only small linear-algebra overhead. Experiments on 2D checker distributions, scaffold decoration, molecular property optimization, and class-conditional CIFAR-10 generation demonstrate recovery from harmful fitted updates. Across molecular and image domains, RCPT mitigates harmful fitted updates and improves a broad range of terminal metrics relative to uncalibrated VCG.
Related
- Multifidelity Computer Model Emulation Via Diffusion Model Steering and Targeted Maximum Likelihood
- Steering Diffusion Models to Rare Events with Sequential Monte Carlo
- FluxLite: Inference-Time Proposal Control for Discrete Diffusion Models
- DASH: Dual-Branch Score Distillation for Guidance-Calibrated Compact Diffusion Models
- Nonparametric Estimation of the Diffusion-Interaction Function in Particle Systems
BibTeX
@misc{lee2026riskcalibrated,
title = {{Risk-Calibrated Proposal Transport for Finite-Particle Diffusion Steering}},
author = {Ziseok Lee and Jaehyeon Kim and Seungwon Kim and Seunghyun Moon and Haneul Choi and Wooyeol Lee and Donghyun Koh and Minhyeong Lee and Kyungsu Kim},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.04171},
eprint = {2610.04171},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.04171},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Risk-Calibrated Proposal Transport for Finite-Particle Diffusion Steering
AU - Ziseok Lee
AU - Jaehyeon Kim
AU - Seungwon Kim
AU - Seunghyun Moon
AU - Haneul Choi
AU - Wooyeol Lee
AU - Donghyun Koh
AU - Minhyeong Lee
AU - Kyungsu Kim
PY - 2026
DA - 2026/10/03
PB - arXiv
DO - 10.48550/arxiv.2610.04171
UR - https://arxiv.org/abs/2610.04171
AB - Inference-time steering combines pretrained diffusion experts or rewards without retraining by changing the dynamics that transport noise to data. Feynman-Kac correction compensates for proposal mismatch through importance-weighted sequential Monte Carlo (SMC), whose finite-particle behavior depends on the proposal. Variance-controlling guidance (VCG) improves that proposal by fitting a linear drift correction to minimize empirical log-weight-rate variance. Although its population optimum cannot worsen residual variance, finite-particle VCG can nearly eliminate its fitting residual while increasing residual risk on new states by orders of magnitude. The resulting update can degrade unweighted generation or accelerate particle collapse. We show that the centered Feynman-Kac rate is the normalized transport residual and that expected out-of-fit benefit is exactly population headroom minus coefficient-estimation penalty. Under regularity assumptions, a Wasserstein analysis bounds the unweighted proposal's terminal error using this residual. These results motivate Risk-Calibrated Proposal Transport (RCPT), which uses deletion leave-one-out residuals to calibrate the retained fraction of the VCG update, adding no model calls and only small linear-algebra overhead. Experiments on 2D checker distributions, scaffold decoration, molecular property optimization, and class-conditional CIFAR-10 generation demonstrate recovery from harmful fitted updates. Across molecular and image domains, RCPT mitigates harmful fitted updates and improves a broad range of terminal metrics relative to uncalibrated VCG.
ER -