Reproduce “ReMaRL: Learning to Unmask and Remask in Frozen Diffusion Language Models” 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
Egor Shibaev, Danila Biktimirov, Ivan Klimov, Artem Makoian, Viacheslav Meshchaninov, Dmitry Vetrov
ReMaRL learns a small policy for a frozen diffusion language model to choose which answer positions to reveal, retain, or mask again. The roughly 0.3-million-parameter policy uses final-answer correctness rewards while the backbone stays frozen. At eight calls on GSM8K, it reaches 53.60% accuracy versus 42.30% for tuned P2′. This supports accuracy gains at tested call budgets, not wall-clock speedups.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.