Ziyang Huang

Trace-Based On-Policy Distillation for Masked Diffusion Language Models

Haolin Ren, Ziyang Huang, Chenhao Yuan, Jun Zhao, Kang Liu

Diffusion large language models (dLLMs) are a promising alternative to autoregressive generation. However, reasoning-oriented post-training for dLLMs remains challenging. Supervised fine-tuning (SFT) for dLLMs requires dense but often off-policy masked states, while reinforcement…Read more