TL;DR of the paper, generated automatically.

New paper: a remasking trick that helps diffusion language models on hard problems.

• Method: a fast sampler for easy questions, frequency-aware remasking for hard ones.
• Idea: favors tokens whose predictions keep changing, to explore other reasoning paths.
• Results: MATH-500 accuracy gains up to 3.73 (LLaDA), 7.33 points (Dream).