Researchers describe Random Remask-Denoise (R²D), a test-time method that sometimes masks generated tokens again before denoising them. The paper says this lets a model revisit decoding errors without fine-tuning or reward feedback. The authors report more than 97% accuracy on hard Sudoku puzzles and more than 99% on 3-SAT; they also say the method beat iterative-correction baselines and LLaDA2.1-mini’s official Q-mode accuracy on HumanEval+. These are results reported in the paper’s abstract, and it does not provide independent validation there.
Why it matters
Researchers working on masked diffusion models can investigate whether remasking improves generated code and logic answers, but the abstract does not establish how well the results generalize beyond the reported tests.
