Report
A tiny recursive reasoner is claimed to improve diffusion models' rule-following
A post says adding a tiny recursive reasoner before each denoising step raised its Sudoku-Extreme result from 2.9% to 71.2%.
TLDR
A post claims bigger diffusion models don't follow rules better, but adding a tiny recursive reasoner (TRM) that thinks before each denoising step does. It lists changes from 2.9% to 71.2% for Sudoku-Extreme, 19% to 75% for mazes and 62% to 89% for CLEVR layouts.
Combined views
42
1 Source, first seen ago
2 reposts
