Reproduce “Random Remasking Scales Masked Diffusion Inference” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhaoyi Zhou, Frank Wu, Andrea Zanette
R²D makes masked diffusion models (MDMs) revisit completed predictions by randomly remasking tokens and denoising them again, rather than freezing every revealed token. With standard-trained models, accuracy rises from 19.7% to 93.5% on Sudoku-Hard and from 86.0% to 98.5% on 3-SAT as correction budgets increase. TCT-trained models reach 97.5% and 99.3%, respectively, on these two logical tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.