GRAM accepted to NeurIPS 2026: recursive reasoning as stochastic latent trajectories, 10M params
haydendevs · x · 2026-10-16
GRAM (Generative Recursive reAsoning Models) has been accepted to NeurIPS 2026, with code, an updated arXiv paper, and a project page released.
Core idea: recursive reasoning models like HRM, TRM, and Looped Transformers are deterministic — same input, same reasoning every time — collapsing plausible reasoning paths into a single attractor. GRAM turns recursion into a stochastic latent trajectory, enabling multiple hypotheses, alternative strategies, and inference-time scaling by width (parallel trajectory sampling) as well as depth.
Notably, the same formulation giving conditional reasoning p(y|x) also makes GRAM a general generative model p(x). The authors report results on tasks like Sudoku-Extreme with only 10M parameters.
More from Research
- Stanford paper: generative representation learning cuts tokenizer training compute by 7-8x — HirokatuKataoka · 2026-10-16
- Cohere Labs to present confidence estimation for reasoning models at EMNLP 2026 — Cohere_Labs · 2026-10-16
- Vals RSI Index extrapolates Google models hitting 130% autonomous AI R&D by Oct 2027 — HaydnBelfield · 2026-10-16
- SpatialClaw Pulled Into Agentic Vision Data; Filtering and Evals Were the Real Lift — CMHungSteven · 2026-10-16
- Visual Lean: drag-and-drop GUI for the Natural Numbers Game runs Lean 4 in WASM — MilesCranmer · 2026-10-16
- Fine-Tuned 0.6B Router Beats Its 27B Sibling: Small Specialized Models Win — mshubham · 2026-10-16