Reproduce “Efficient and Unbiased Sampling of Boltzmann Distributions via Consistency Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fengzhe Zhang, Jiajun He, Laurence I. Midgley, Javier Antorán, José Miguel Hernández-Lobato
The authors introduce a diffusion-based importance-sampling method for estimating properties under Boltzmann distributions. It combines learned trajectory steps with Gaussian noise, making transition densities available for weighting, and designs a matching reverse-direction target path. On GMM-40 in 2D, BCTM with importance sampling reports 2.8 ± 1.3% ESS using 12 NFEs, versus 2.6 ± 1.1% with 100 NFEs for DDPM; this compares evaluation budgets, not wall-clock time or exact finite-sample unbiasedness.
Key insights
- Fast generative model distillation (via Consistency Models) can be successfully integrated with statistical bias correction (via Importance Sampling) by introducing controlled stochastic steps.
- Constructing both the proposal and target distributions using a similar alternating ODE-SDE structure significantly enhances their alignment and the efficiency of Importance Sampling.
- Equivariance in consistency models is better handled by extending Consistency Trajectory Models (CTMs) to Bidirectional CTMs (BCTMs), especially for molecular systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.