GRAM paper code released: recursive reasoning models with Bengio as co-author
SungjinAhn_ · x · 2026-10-16
The official codebase for Generative Recursive Reasoning (GRAM) is now public under MIT license, with the paper submitted to NeurIPS 2026. Authors include Junyeob Baek, Mengye Ren, and Yoshua Bengio.
The paper explores recursive reasoning models (RRMs) as an alternative to autoregressive sequence extension: instead of generating longer chains, RRMs iteratively refine a latent state via shared transition functions. GRAM addresses the limitation that existing RRMs are largely deterministic, following a single latent trajectory. The repo ships pretraining scripts, datasets, evaluators, and pass@k evaluation code.
More from Research
- FreeSolo: Linear-Time Sparse Attention With Zero-Shot Length Extrapolation — Zekun Shi · 2026-10-16
- Apollo's alignment-faking experiment: model fakes compliance on 12% of responses when training is at stake — sebkrier · 2026-10-16
- VPO: The Post-Training Method That Teaches Models to Output Diverse Good Answers — lateinteraction · 2026-10-16
- 797GB of n-grams in read-only CPU RAM match 57% more compute at 1B-scale LM pretraining — anshulkundaje · 2026-10-16
- New essay: prioritizing falsifiability in AI cognitive modeling can be reward-hacked — AndrewLampinen · 2026-10-16
- SPLICE: repairing failed math reasoning with solution hints boosts AIME and proofs — behrouz_ali · 2026-10-16