Mask-Aware Policy Gradients for Diffusion Language Models
The paper treats MDLM generation as two linked choices: which tokens to fill in, and which positions to mask again.
The authors argue that reinforcement learning for masked diffusion language models has been limited by hard-to-estimate log likelihoods. Their method decomposes the policy gradient into a token term and a masking term, rather than optimizing token predictions alone. They report state-of-the-art results on math reasoning and coding benchmarks, including 87.1% on GSM8K and 53.4% on MBPP. The paper was accepted at COLM 2026. ArXiv · AI/CL/LG's note
The authors argue that reinforcement learning for masked diffusion language models has been limited by hard-to-estimate log likelihoods. Their method decomposes the policy gradient into a token term and a masking term, rather than optimizing token predictions alone. They report state-of-the-art results on math reasoning and coding benchmarks, including 87.1% on GSM8K and 53.4% on MBPP. The paper was accepted at COLM 2026. ArXiv · AI/CL/LG's note
score 5