Reproduce “Train for Many, Update with One: Improving Multi-Sample Reasoning via Fenchel-Bregman Posterior Learning” 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

Sungjun Lim, Adrian Robert Minut, Nico Carlsen Daheim, Mohammad Emtiyaz Khan, Kyungwoo Song, Thomas Möllenhoff

FBI trains a solver-parameter distribution to improve the chance that at least one of K sampled candidates is correct. It reformulates the finite-K all-fail objective with problem-specific weights, then uses one posterior draw per update while tracking those weights. On Maze-Hard, FBI reaches 91.3% Pass@10, versus Direct-10’s 91.0%, with lower measured time per update, not necessarily lower total training time.

Summary by alphaXiv.

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