Research
Research proposes DTCA, a token credit assignment method for RL in diffusion language models
DTCA weights tokens by how much they raise prediction confidence of remaining masked positions and adjusts signals by entropy changes. On LLaDA-8B-Instruct it reports gains of up to 9.0 points on MATH500 and 5.3 points on GSM8K over the base model.
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
DTCA: Denoising Token Credit Assignment for Reinforcement Learning in Diffusion Language Models