Reproduce “WASABI: Wasserstein Sampling via Bridge Interpolants” 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

Sebastian Salazar, Dilum Fernando, Michal Kucer, Emily Casleton, Luhuan Wu, David M. Blei

WASABI trains a diffusion sampler through Wasserstein descent on reverse KL, using score-free regression to avoid requiring the model’s log-density gradient. After training, it generates samples without further molecular-potential evaluations. On alanine dipeptide at 300 K, it reports a Jensen–Shannon divergence of 0.0254 ± 0.0033 nats, versus 0.068 ± 0.009 for the best reported baseline; its 3.13 million potential-call budget is estimated, not an elapsed-time or total-computation comparison.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.