Diffusion LMs Bring Freedom to the Order of Thought

furongh · x · 2026-07-06

Autoregressive language models inherently generate tokens sequentially from left to right (token1→token2→token3…), making generation and reasoning a fundamentally serial process. Diffusion language models differ: they start with masked text and iteratively reveal tokens, introducing a new kind of freedom—the order in which thoughts emerge.

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

Original post →

More from Research

Research channel →