arxivcs.LGcs.AI2026-07-20
RRPO: Reference-Relative Policy Optimization with Stratified Conditional Rollouts
Yuxin Xiong, Xunyi Jiang, Rohan Surana, Xintong Li, Sheldon Yu, Nikki Lijing Kuang, et al.
Group Relative Policy Optimization (GRPO) has shown strong effectiveness in reinforcement learning from verifiable feedback, where sampled rollouts can be compared within a group using task-provided correctness signals. However, extending group-relative optimization beyond verifi…