Simple Approximation and Derivative Free Inference-Time Scaling for Diffusion Models via Sequential Monte Carlo on Path Measures
Published in International Conference on Machine Learning (ICML), 2026
URGE attaches a Girsanov path weight to every simulated trajectory and periodically resamples: gradient-free inference-time scaling with no score, Hessian or PDE evaluations, and the same terminal law as particle-wise SMC.
Recommended citation: Chenyang Wang*, Weizhong Wang*, Yinuo Ren, Jose Blanchet, and Yiping Lu. "Simple Approximation and Derivative Free Inference-Time Scaling for Diffusion Models via Sequential Monte Carlo on Path Measures." In Proceedings of the 43rd International Conference on Machine Learning (ICML), 2026. (* equal contribution)
Download Paper | Download Bibtex