From Correctness to Preference: A Framework for Personalized Agentic Reinforcement Learning

In three linesPARPO framework for personalized agentic reinforcement learning. Decouples generic task rewards from user-specific preference rewards using user anchors. Introduces PSGM for preference-aligned skill retrieval. Evaluated on ETAPP, ETAPP-Hard, SJAgent with improvements over memory and RL baselines.

Summary generated by Claude — human-verified