Ruizhi Qiao

RSPO: Reward-Swap Policy Optimization for Multi-Turn LLM Agents

Qiang Liu, Taian Guo, Ruizhi Qiao, Xing Sun

Reinforcement learning holds significant potential for training large language models (LLMs) to handle multi-turn interactive tasks. However, in long-horizon, multi-turn tasks characterized by sparse outcome rewards, directly training with outcome rewards often results in slow co…Read more