Large language models have moved beyond single-turn chatbots into interactive agents that must reason over long horizons. In real deployments, however, MCP tool services often return noisy or erroneous feedback, and users may interrupt to correct the agent. Existing methods offer limited supervision for learning reflection and recovering from failure.
To address this, researchers propose Agent-Reflex, a general training framework designed to incentivize proactive reflection in agents. The framework has three stages. First, Environment Reflection Skill Discovery (RSD) injects grounded tool-feedback perturbations into environment scaling, samples a capability-diverse LLM pool, mines matched success-failure pairs, and summarizes reusable reflection meta-skills.
Second, Reflective Supervised Fine-Tuning (ReSFT) converts each discovered skill into supervision for self-reflection, skill recovery, and user-centric correction, strengthening complementary reflective behaviors. Third, On-Policy Reflective Policy Optimization (ORPO) combines agentic reinforcement learning under noisy interactions with reflection-guided on-policy distillation to further elicit proactive reflection.
Built on Qwen3-8B, Agent-Reflex reaches 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, surpassing strong environment-scaling and reflection baselines. The authors report that RSD injects real tool failures into 557 MCP environments and distills 1,170 reflection meta-skills from success/failure contrasts.
Experiments across six challenging agent benchmark suites demonstrate the framework's effectiveness. Further analyses show Agent-Reflex improves robustness to noise while maintaining stable training dynamics, supporting scalable agent training. Code is available at github.com/RUC-NLPIR/Agent-Reflex.