Persistent-memory denoiser hits 99.9% on Sudoku-Extreme with under 250k parameters

LucaAmb · x · 2026-09-12

Researchers added a persistent memory state and removed timestep conditioning from a denoising model, yielding an anytime iterative solver. Since classic sudoku solvers need memory to track possible states, persistent memory between denoising steps proves key: with fewer than 250k parameters the model scores 99.9% on Sudoku-Extreme and 98.3% on Maze-Unique — a new take on reasoning with atypical diffusion models.

Related event: Denoiser with Persistent Memory Solves Sudoku-Extreme at 99.9%(2 posts)→

Original post →

More from Research

Research channel →