Autoregressive Boltzmann Generators sample molecular equilibria with a transformer

Category: Paper
Status: developing
Sources: arXiv, code
Summary: Rehman, Tan, Bengio, Bose, and Tong published Autoregressive Boltzmann Generators (ArBG), an ICML 2026 spotlight, on 2026-06-25. The method replaces the normalizing-flow backbone used in prior Boltzmann generators with an autoregressive transformer plus sequential inference-time interventions, removing the invertibility constraint that limits flow-based equilibrium sampling of molecular systems. The authors introduce Robin, a 132M-parameter transferable model, and report cutting the zero-shot energy error (E-W2) on 8-residue peptide systems by over 60%. The numbers are the authors' own; code is released.
Why it matters: Transferable equilibrium sampling with an LLM-style architecture lowers the cost of generating molecular conformations, relevant to molecular dynamics and drug-discovery pipelines.
Follow-up: Watch for independent reproduction of the 8-residue energy-error result, transfer to larger peptides or proteins, and adoption of Robin as a pretrained sampler.