Efficient Boltzmann Sampling
Consistency models with importance sampling for efficient, unbiased Boltzmann distributions.
This project studies how consistency models can be used as fast proposal samplers for Boltzmann distributions while importance sampling corrects the remaining bias.
The core result is a reduction in function evaluations from 100 to 6-25 while preserving effective sample size on synthetic and equivariant n-body systems.
Related publication: Efficient and Unbiased Sampling of Boltzmann Distributions via Consistency Models.