Probe guidance steers continuous diffusion LMs with just 1-3% extra inference compute

itsbautistam · x · 2026-09-19

rohitdilip highlights probe guidance for continuous diffusion language models: good results with only a 1-3% increase in inference compute, and it's domain agnostic. The authors are excited by recent progress on continuous dLMs but note much remains open on how best to guide language models.

Related event: Apple Proposes Probe Guidance for Diffusion Language Models(2 posts)→

Original post →

More from Research

Research channel →