What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models

A controlled study of learned unmasking policies in masked diffusion language models finds that the policies, which decide which token predictions to commit during parallel generation, improve the speed-accuracy trade-off but vary in supervision, inputs, and capacity, leaving unclear which design choices drive performance. The research frames supervised policies as the basis for isolating those variables.

Masked diffusion language models generate tokens in parallel, but their efficiency depends on which predictions are committed. Learned unmasking policies improve the speed--accuracy trade-off, yet differ in supervision, inputs, and capacity, obscuring which choices matter. We view supervised policie