Diffusion LM Decoding Doesn't Have to Be Left-to-Right

furongh · x · 2026-07-06

In masked diffusion LMs, decoding doesn't have to occur strictly from left to right; the model can choose the next masked position to reveal. In principle, this opens the door to more flexible and ultimately more parallel reasoning, though it presents a critical challenge.

Related event: ICML 2026 Paper SAS: Optimizing Thought Scheduling in Diffusion LMs(15 posts)→

Original post →

More from Research

Research channel →