T2M Refinement for Diffusion Language Model Consistency

Token-to-Mask (T2M) is a training-free inference-time method that improves diffusion language models by remasking low-confidence tokens during the denoising process. This correction mechanism transfers across tasks and models without requiring retuning.