Time Scaling via Stochastic Simulation: Guidance, Correction and Sequential Monte Carlo

Data Science Seminar

Yiping Lu
Northwestern University

Abstract

Inference-time scaling is the practice of improving a model’s output quality by spending more computation at inference, such as using longer reasoning, broader search, or multiple candidate evaluations. In this talk, we introduce a stochastic simulation framework for designing inference-time scaling strategies across diffusion models, language model reasoning, and high-dimensional PDE solvers. The first component leverages a value function that predicts future generation quality to guide the generation process. The second introduces a correction step, based on sequential Monte Carlo, to compensate for potentially suboptimal value-function estimates. I will discuss both new inference-time scaling algorithms, share our estimation results, and outline open questions on sample complexity.