Report
25M-parameter model claimed to solve 86.1% of Hard Sudoku boards in 12 passes
A post puts greedy autoregressive decoding at 1.0%, diffusion model DUO at 58.4% and flow-map model FMLM+ at 81.3%.
TLDR
A user reports that a 25M-parameter non-autoregressive model solved 86.1% of Hard Sudoku boards in 12 forward passes. Their comparison puts greedy autoregressive decoding at 1.0%, discrete diffusion model DUO at 58.4% in 128 passes and flow-map model FMLM+ at 81.3% in 16.
Combined views
—
1 Source, first seen ago
