Researchers introduce Group Adaptive Clipping Policy Optimization (GAPO), a modification to Group Relative Policy Optimization methods used in reinforcement learning with verifiable rewards. The method addresses a limitation in which rare correct solutions to hard problems get clipped at the same rate as abundant correct solutions to easy problems, despite carrying much stronger learning signal, by adapting the clipping boundary to each rollout’s advantage. Across Qwen and Llama models, GAPO consistently improves both Pass@1 and Pass@k accuracy over fixed-clipping and advantage-shaping baselines on math reasoning and coding benchmarks.
