Mask-Aware Policy Gradients for Diffusion Language Models

arXiv cs.CL · 6 hours ago

Researchers developed a reinforcement learning approach for Masked Diffusion Language Models that accounts for both token selection and the order of position unmasking during generation. The method achieved 87.1% on GSM8K and 53.4% on MBPP mathematical reasoning and coding benchmarks. This approach enables RL-based improvement of reasoning capabilities in MDLMs by properly modeling the full generation process rather than just token predictions.