Variational Inference Meets Sampling

← Back to Projects

This ongoing project explores the interface between variational inference and sampling-based methods (MCMC, Stein Variational Gradient Descent, diffusion-based samplers). The goal is to design algorithms that combine the speed and amortization of VI with the asymptotic correctness and multi-modality of sampling.

Topics we are currently investigating include:

• Scalable entropy estimation for unnormalized distributions via SVGD dynamics.
• Energy-based reinforcement learning with Stein-style policies (see S2AC, ICLR 2024).
• Connections between score-based diffusion samplers and amortized variational methods.

Relevant publications:

• Particles Don’t Care About Z: Towards Scaling Entropy Estimation of Unnormalized Densities — ICML 2026.
• S2AC: Energy-Based Reinforcement Learning with Stein Soft Actor Critic — ICLR 2024.