Sakana AI· Sakana modelsResearch paper

UnMaskFork: Test-Time Scaling for Masked Diffusion via Deterministic Action Branching

Sakana AI announces its ICML 2026 paper 'UnMaskFork' which introduces a method for test-time scaling in masked diffusion language models by having multiple models collaborate through model switching and Monte Carlo Tree Search, improving performance on coding and math tasks without additional training or model changes.

From the source

In our ICML 2026 paper “UnMaskFork,” we show that having multiple masked diffusion language models collaborate on a single answer improves performance on coding and math tasks.

sakana.ai

Announced

22 Jul 2026 15:00 UTC

Detected

28 Aug 2026 21:48 UTC (+37d after)