Agent-Reflex trains language agents to reflect on noisy tool feedback
A new arXiv paper, Agent-Reflex, proposes a training framework to teach language agents to reflect when MCP tool services return noisy feedback or users step in to correct them. It has three stages: Environment Reflection Skill Discovery mines matched success-failure pairs into reusable reflection meta-skills, Reflective Supervised Fine-Tuning converts those skills into supervision for self-reflection and recovery, and On-Policy Reflective Policy Optimization combines agentic reinforcement learning under noisy interactions with reflection-guided on-policy distillation. Built on Qwen3-8B, the authors report 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, saying it surpasses environment-scaling and reflection baselines; the results are author-reported and not independently verified, and the paper says its code is available.
The authors argue that noisy tool feedback and user corrections are common in real agent deployments, while existing methods provide limited supervision for reflection learning and failure recovery.
Computer Science > Computation and Language
arXiv:2610.17953 (cs)
[Submitted on 13 Oct 2026]
Title:Agent-Reflex: Teaching Language Agents to Act Reflectively
Authors:Guanting Dong, Xinyu Yang, Jiajie Jin, Yuyang Hu, Xiaoshuai Song, Yifei Chen, Xiaoxi Li, Jiejun Tan, Yutao Zhu, Tat-Seng Chua, Ji-Rong Wen, Zhicheng Dou
View a PDF of the paper titled Agent-Reflex: Teaching Language Agents to Act Reflectively, by Guanting Dong and 11 other authors