Unmask the State: When Does State Adaptation Matter for Masked Diffusion Language Models
The paper argues that masked diffusion models should change decoding strategy only in selected high-opportunity states.
It frames inference choices across score, cardinality, region, commitment, and planning, then tests when an alternative action beats a validation-picked fixed action. Across three MDMs and ten tasks, those “strategy reversals” are uneven rather than constant. A lightweight detector captured 56.9% of the candidate-set oracle opportunity on LLaDA-8B constrained JSON filling while adapting only the top 10% of states. Source: HF Daily Papers' note.
It frames inference choices across score, cardinality, region, commitment, and planning, then tests when an alternative action beats a validation-picked fixed action. Across three MDMs and ten tasks, those “strategy reversals” are uneven rather than constant. A lightweight detector captured 56.9% of the candidate-set oracle opportunity on LLaDA-8B constrained JSON filling while adapting only the top 10% of states. Source: HF Daily Papers' note.
score 4