Decoding Order Is Itself a Planning Problem

nandofioretto · x · 2026-07-17

An ICML 2025 paper on masked diffusion reveals an interesting finding: while models learn all token infilling orders during training, you can decide which position to unmask first during inference.

The authors note that by simply choosing the right decoding order, Sudoku task accuracy can jump from roughly 6% to 89%. This demonstrates that decoding order is fundamentally a planning problem, rather than a trivial implementation detail.

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

Original post →

More from Research

Research channel →