WASABI: Blei lab trains diffusion models to sample Boltzmann distributions with 3.3x fewer energy evaluations
skoularidou · x · 2026-11-02
David Blei's Columbia group introduced WASABI (Wasserstein Sampling via Bridge Interpolants), addressing the fundamental problem of sampling from Boltzmann distributions in physics, chemistry, and biology.
- Trains a diffusion model to generate samples from a target Boltzmann distribution given only the energy function, minimizing KL divergence via Wasserstein gradient flows — building on principles behind Langevin dynamics
- Key hurdle: applying Wasserstein gradient flows to diffusion models requires the model's own log-density gradient (intractable); the authors derive a practical training algorithm avoiding it
- After training, sampling needs no further energy evaluations; on alanine dipeptide it beats the best reported baselines with 3.3x fewer energy evaluations, and is competitive on interacting particle system benchmarks
More from Research
- Essay: whether AI ends or elevates human math depends on two models of progress — Dr_Atoosa · 2026-11-02
- Qwen 27B orchestrates SAM 3 per tile: one H200 serves 12 patients in medical imaging — MaziyarPanahi · 2026-11-02
- Rentosertib: a 7-year quest for an aging drug worth 110 million human lifetimes — rand_longevity · 2026-11-02
- GeBNLP 2026 goes virtual Nov 9 at AACL with keynotes from Savoldi and Gautam — debora_nozza · 2026-11-02
- From-scratch 128x128 rectified-flow model draws newsprint faces on CPU in 0.27s, MIT licensed — rustydustyshckleford · 2026-11-02
- Berkeley's Evolving Agent Teams beats production FlashInfer by 1.2x on CUDA kernels from scratch — berkeley_ai · 2026-11-02