Steering Masked Diffusion via Token Commitment Control
November 18, 2026
This method uses token commitment as a control surface for masked diffusion models like LLaDA. It introduces counterfactual target-aware commitment and adaptive budgets to improve target quality in language and vision-language generation.
HOW THIS AFFECTS YOU
●
researcherYou can steer masked diffusion models by controlling which and how many tokens are committed during sampling.