Reproduce “DTCA: Denoising Token Credit Assignment for Reinforcement Learning in 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

Hongze Chen, Miao Pan, Lizhou Liang, Qinfeng Li, Xinyin Ma, Yongliang Shen, Weishi Wang, Daniel Dahlmeier and 2 more

DTCA improves credit assignment for diffusion language models by estimating how committed tokens change confidence at masked positions. It averages estimates by part of speech, then uses weights to select denoising steps and assign reinforcement-learning credit. On LLaDA-8B-Instruct, it reaches 39.8% MATH500 accuracy versus 30.8% for the base model at 128 tokens; tests cover one model and mathematical reasoning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.