Decoding Order Determines Sudoku Success

nandofioretto · x · 2026-07-17

This post discusses an outstanding masked diffusion paper from ICML 2025. The core finding is that while models train on "fill-in-the-blank orders for all tokens," they can freely decide which position to solve next during inference. The author highlights a fascinating result: simply by choosing the right decoding order, accuracy on Sudoku tasks can jump from roughly 6% to about 89%. This shows that in such models, **the decoding order itself is a planning problem**, not just a minor execution detail.

Related event: ICML Paper: Decoding Order as Planning in Diffusion(2 posts)→

Original post →

More from Research

Research channel →