Mariia Drozdova Describes Anytime Iterative Solver
PhD student at University of Geneva posts results from modified denoising model.
TLDR
Mariia Drozdova posted a thread describing changes to a denoising model. She added a persistent memory state and removed timestep conditioning. The post states this produces an anytime iterative solver. Drozdova reported the model has under 250k parameters and reaches 99.9 percent accuracy on Sudoku-Extreme plus 98.3 percent on Maze-Unique. The tweet links to an arXiv paper and includes a side-by-side video animation of the two tasks.
Combined views
37.1K
1 Source, first seen 27d ago