Reproduce “Generative Recursive Reasoning” 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

Junyeob Baek, Mingyu Jo, Minsu Kim, Mengye Ren, Yoshua Bengio, Sungjin Ahn

GRAM lets recursive reasoning explore multiple latent paths by learning state-dependent stochastic transitions: each adds Gaussian noise to a deterministic update. On Sudoku-Extreme, 20 parallel samples at 16 iterations reach 96.06% accuracy in 184 ms, versus TRM’s 92.15% at 320 sequential iterations in 2,755 ms, with nearly matched compute; this is a reported single-request result, not reduced computation.

Summary by alphaXiv.

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