PGPO: Potential-Guided Policy Optimization for Multi-Turn Agentic Tasks
Proposed Potential-Guided Policy Optimization (PGPO) provides finer-grained step-level credit assignment, especially within failed trajectories, and shows strong overall performance relative to recent group-based RL methods.
Yu-Yao Zheng, Haipeng Sun, Jun-Wei Bao et al.
· 1 citation