一作ACL 2026Poster· Natural Language Processing
d- T ree RPO : Towards More Reliable Policy Optimization for Diffusion Language Models
Leyi Pan · Shuchang Tao · Yunpeng Zhai · Zheyu Fu · Liancheng Fang · Minghua He · Lingzhe Zhang · Zhaoyang Liu · Bolin Ding · Aiwei Liu · Lijie Wen
摘要
Reinforcement learning (RL) is pivotal for enhancing the reasoning capabilities of diffusion large language models (dLLMs). However, existing dLLM policy optimization methods suffer from two critical reliability bottlenecks: (1) reward sparsity, arising from coarse or unverifiable signals that im...