Token-to-Mask Remasking Improves Discrete Diffusion Language Models

A new training-free method called Token-to-Mask (T2M) remasking addresses limitations in discrete masked diffusion language models like LLaDA. T2M replaces the Token-to-Token (T2T) editing mechanism introduced in LLaDA2.1, which directly replaces suspected erroneous tokens. T2M resets such tokens back to the mask state, enabling the diffusion process to re-predict them under a cleaner context. The approach decouples error detection from replacement, avoids polluting the generation context, and eliminates the train-inference noise mismatch caused by systematic model-generated errors. The authors design and empirically validate three complementary error detection strategies. The paper is available on arXiv under identifier 2605.26436.