Paper page - Group Adaptive Clipping Policy Optimization

Sheng Jia and others proposed a new method called “Group Adaptive裁剪 Strategy Optimization (GAPO)” to improve group relative strategy optimization in reinforcement learning. This method addresses the problem of over-suppression of relevant samples caused by existing fixed importance sampling裁剪 boundaries. It solves this problem by dynamically adjusting the boundaries based on rolling advantages. GAPO does not require modifications to the reward shaping or objective function, and improved the mathematical reasoning benchmarks in pass@1 and pass@k on both Qwen and Llama-based models.