Tiny Recursive Reasoner Boosts Diffusion Models: Sudoku 2.9% to 71.2%

jm_alexia · x · 2026-10-14

Bigger diffusion models don't follow rules better — but adding a tiny recursive reasoner (TRM) that thinks before each denoising step does: Sudoku-Extreme jumps from 2.9% to 71.2%, Mazes from 19% to 75%, and CLEVR layouts from 62% to 89%.

Original post →

More from Multimodal

Multimodal channel →