Generative Recursive Reasoning

Summary

GRAM adds learned stochastic residual guidance ε~N(μθ,σθ²I) on the high-level latent of an HRM/TRM-style hierarchy and trains via a truncated ELBO. Width scaling samples N trajectories and ranks them with an LPRM or majority vote. Sudoku-Extreme 97.0% vs TRM 87.4% at 16 steps; N=20 at 16 iters beats TRM at 320 (97.0 vs 90.5). ARC-1/2 52.0/11.1 vs TRM 44.6/7.8. N-Queens 8×8 accuracy 99.7% / coverage 90.3% vs TRM 66.8/36.1. Unconditional Sudoku 99.05% validity (10.9M, 16 steps) vs D3PM-Big 91.33% (55.1M, 1000 steps); binarized MNIST FID 73.34 at 256 steps vs TRM collapse 303. Project https://ahn-ml.github.io/gram-website/.

Keywords

gram · recursive-reasoning · trm · hrm · variational-inference · sudoku · arc-agi · n-queens · kaist · mila · bengio

Topics

recursive reasoning, latent-variable models, generative modeling

Research notes

• Primary: arxiv abs (cs.AI). CC BY 4.0 on HTML. KAIST/Mila/NYU/UdeM. Baek/Jo/Kim equal contrib. Correspondence Junyeob Baek wnsdlqjtm@kaist.ac.kr, Mingyu Jo mingyu.jo@kaist.ac.kr, Sungjin Ahn sungjin.ahn@kaist.ac.kr. Project page says code coming soon. HF paper page 31 upvotes; no linked models/datasets. Discord posted abs (also AlphaXiv). Uses public Sudoku-Extreme/ARC-AGI/MNIST plus author-generated N-Queens/Graph Coloring; no standalone public corpus, so no datasets_local row. License field left blank per catalog convention.