Multi-Turn Reflective Masking Elicits Reasoning in Mask Diffusion Models

Reflective Masking (RM) enables multi-turn iterative refinement in Mask Diffusion Models by selectively re-masking and denoising only the tokens requiring correction, rather than regenerating full sequences. Trained via lightweight post-training, RM gives MDMs native test-time scaling without the sequential dependency of autoregressive chain-of-thought approaches.