Philipp Krähenbühl

Mask-Aware Policy Gradients for Diffusion Language Models

Haran Raajesh, Kulin Shah, Adam Klivans, Philipp Krähenbühl

Reinforcement learning has proven effective for improving reasoning in large language models, but extending it to Masked Diffusion Language Models (MDLMs) remains challenging due to the intractability of the log-likelihood estimation. Existing approaches approximate this log-like…Read more