Multi-Turn Reflective Masking Elicits Reasoning in Mask Diffusion Models
Researchers have introduced Reflective Masking (RM), a novel post-training technique designed to enhance reasoning capabilities in Mask Diffusion Models (MDMs). Unlike autoregressive models that rely on sequential generation, RM allows MDMs to iteratively refine previous outputs through multi-turn masking and denoising, mirroring human error correction. The method incorporates History Reference, a parameter-free mechanism that utilizes intermediate denoising states to leverage insights from prior turns. This approach requires no architectural changes and has demonstrated consistent performance improvements across various tasks, including text generation, Sudoku, and image editing, positioning RM as a foundational element for MDM reasoning. AI
IMPACT This research could enable diffusion models to perform more complex reasoning tasks, potentially improving their utility in areas like creative generation and problem-solving.