Sakana AI’s UnMaskFork uses multiple diffusion LMs to scale inference at test time

SakanaAILabs · x · 2026-07-29

Sakana AI says its ICML-covered work on UnMaskFork explores how to use the collective intelligence of multiple diffusion language models to enable inference-time scaling.

According to the cited Nikkei xTECH piece, the team argues that inference-time scaling will remain useful even as models improve, because the method can keep applying to future generations of diffusion language models. The research is still early, but the article frames it as a foundational theory effort for a still-maturing field.

Original post →

More from Research

Research channel →