Agent-Reflex Lifts Qwen3-8B to 58.6 on Tool-Use Benchmarks

A working-draft preprint posted to arXiv by a Renmin University-led team reports that Agent-Reflex, a reflection-training framework built on Qwen3-8B, reaches 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench.

The paper's framing is blunt: '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.' Agent-Reflex answers that gap in three stages. Environment Reflection Skill Discovery '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.' Reflective Supervised Fine-Tuning turns those skills into training signal for self-reflection, skill recovery, and user-centric correction. On-Policy Reflective Policy Optimization then layers agentic reinforcement learning under noisy interactions with reflection-guided on-policy distillation.

The authors mark the work 'in progress' and release code at a RUC-NLPIR GitHub repo. The abstract names six 'challenging agent benchmark suites' used in experiments but does not list them, and publishes no per-benchmark baseline numbers. It lands in a crowded corner of the field: our agents feed has carried 475 stories in the last 90 days.