Team member reports 99.9% on Sudoku Extreme with diffusion models
Giving denoisers—the parts of diffusion models that remove noise—memory that persists between steps seems key, a team member says.
TLDR
A team member reports 99.9% on Sudoku Extreme in research exploring reasoning with atypical diffusion models. They draw a parallel with classic Sudoku solvers, which they say need memory to track possible states, and suggest that persistent memory between denoising steps seems key.
Combined views
—
1 Source, first seen ago
