Agent-Reflex：教语言智能体进行反思性行动
原标题：Paper page - Agent-Reflex: Teaching Language Agents to Act Reflectively
AI 摘要
中国人民大学 RUC-NLPIR 团队提出 Agent-Reflex 训练框架，旨在让语言智能体在噪声工具反馈和用户干预下具备主动反思能力。该框架包含三个阶段：环境反思技能发现（RSD）向 557 个 MCP 环境注入真实工具故障并挖掘 1,170 条反思元技能；反思监督微调（ReSFT）将技能转化为自反思、恢复和用户纠正的监督数据；在线策略反思策略优化（ORPO）结合智能体强化学习与反思引导的在线策略蒸馏。基于 Qwen3-8B，Agent-Reflex 在五个长程工具使用基准上达到 58.6 总分（+18.4），在 AgentNoiseBench 上达到 29.1，超过 Agent-World、EnvScaler 和 NoisyAgent 等基线，代码与模型已开源。
正文节选
Agent-Reflex: Teaching Language Agents to Act Reflectively Abstract Large language models have evolved from single-turn chatbots into interactive agents capable of long-horizon reasoning. In real deployments, MCP tool services can return noisy feedback and users may intervene to correct the agent, but existing methods provide limited supervision for reflection learning and failure recovery. To bridge this gap, we propose Agent-Reflex, a general training framework for incentivizing proactive refl