AI RESEARCH
d-TreeRPO: Towards More Reliable Policy Optimization for Diffusion Language Models
arXiv CS.CL
•
ArXi:2512.09675v3 Announce Type: replace 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 impede accurate advantage calculation; and (2) their probability estimates do not account for the gap to the unbiased expectation over all decoding orders, which are intractable to compute.