Skip to main content
Search

Mingwei Xu

1 related articles

Subscribe to updates
  1. arXiv

    POPO learns only from positive rollouts via online self-distillation, reaching 36.67% on AIME 2025 with Qwen-Math-7B versus GRPO's 30.00%

    The work proposes Positive-Only Policy Optimization (POPO), an on-policy self-distillation framework integrated with RLVR that learns exclusively from online positive rollouts, using bounded importance sampling, a Siamese policy network with a momentum-based adaptation law, and a bounded similarity penalty instead of KL divergence; across MATH-500, AMC23, AIME 2024/2025, and Olympiad it outperforms GRPO, with Qwen-Math-7B reaching 36.67% on AIME 2025 versus GRPO's 30.00%.

You’re all caught up