全部文章0
Hugging Face PapersZhengpeng Xie, Li Lyna Zhang, Zeke Xie, Mao Yang··访问 1

Trust Region Policy Distillation

原网页

论文信息

  • arXiv ID 2607.04751
  • 作者 Zhengpeng Xie, Li Lyna Zhang, Zeke Xie, Mao Yang
  • 链接 arXiv · PDF · Hugging Face

摘要

Big goals are hard to achieve all at once; breaking them into small steps is wiser. We present Trust Region Policy Distillation (TOP-D), which transforms the notoriously unstable, high-variance On-Policy Distillation (OPD) into a stable training paradigm by dynamically constructing a proximal teacher. Theoretically, we establish a rigorous framework demonstrating that TOP-D inherently controls gradient variance. By providing a formal global convergence analysis alongside a monotonic improvement bound, we mathematically formalize the reliability and stability of the overall training dynamics. Empirically, TOP-D dramatically enhances training stability, sample efficiency, and final performance on mathematical reasoning tasks. More importantly, TOP-D introduces zero additional computational overhead, positioning itself as a promising alternative to the well-established OPD paradigm.