Tag
This paper introduces Reflective Masking, a lightweight post-training method that enables mask diffusion models to perform multi-turn self-revision through token-level revision policies and history references, improving performance on reasoning tasks like Sudoku, math, code generation, and image editing.