About

Given the power of diffusion models for the task of generative modeling---generating ``new'' samples given existing ones---it is natural to ask if they have applications to the task of sampling, namely, generating samples given access only to evaluations of an unnormalized density. The sampling problem is most prominent in the field of Bayesian inference in which the density of interest is a posterior distribution in a statistical model and in scientific contexts where it corresponds to the distribution over configurations of some physical system at thermal equilibrium given a description of its Hamiltonian. But this setup also finds more modern applications such as conditional generation given a prompt, or the problem of {\bf reward guidance} in which we aim to better align the output of a large pretrained generative model with a user-specified reward function. In all of these settings, the density from which we desire samples is a "tilt'' of a base density (e.g., the prior density, the uniform distribution over some domain, or the density of the trained model before reward guidance), which leads to the following fundamental algorithmic problem: can we efficiently sample from tilted distributions, given access only to score evaluations (or other information) from the base density? This problem has attracted significant attention in the recent literature, from communities working on LLMs, reinforcement learning, optimal transport, sampling, and stochastic control. The goal of this workshop is to bring together researchers in these fields, together with the TCS community which has long investigated the structure of tilts for theoretical purposes (e.g., Eldan's stochastic localization).

Chairs/Organizers