CoDR Refines Diffusion Language Models via Confidence-Drift Remasking

CoDR is a training-free, sampler-agnostic method that addresses confidence drift in masked diffusion language models. It uses k-partition probing to identify committed tokens that the model no longer endorses in denser context, remasking and regenerating only those specific tokens.