Remask, Don't Replace: Token-to-Mask Refinement in Masked Diffusion Language Models

Implement a 'Remask' refinement loop for masked language models to fix generation errors. This replaces rigid Token-to-Token editing with more intelligent confidence-aware re-masking.

Suggested repo: diffRefine

"Stop replacing, start remasking: improve your generative LLM output quality."

Estimated effort: 50h