Knowledge Editing for Masked Diffusion Language Models
A simple correction is introduced that optimizes the edit including such states in which the target is partially unmasked, substantially restoring multi-token performance in masked diffusion models.
Haewon Park, Yohan Jo
· 1 citation