Multi-turn agents can be trained with terminal rewards, meaning only the final outcome is used as a training signal, avoiding the need for step-by-step annotations. However, the abstract notes that group-relative advantages—a common way to compare outcomes within a batch—provide little guidance about which intermediate decisions should be changed.

The paper introduces PEARL (Process-Experience Aware Reinforcement Learning). The abstract indicates that failed trajectories also contain reasonable elements, suggesting that PEARL leverages process experience from these trajectories to extract more granular training signal. The full mechanism is not detailed in the excerpt.

By addressing the credit-assignment gap, PEARL aims to make terminal-reward training more effective for multi-turn tasks without requiring expensive step-level labels. The paper is available on arXiv under ID 2610.14636.