The Pulse
Agent-Reflex Trains Qwen3-8B to Recover From Tool Failures
A paper on arXiv describes Agent-Reflex, a three-stage training method for helping Qwen3-8B recover from faulty tool responses and user corrections. The authors report scores of 58.6 across five long-horizon tool-use benchmarks and 29.1 on
AI.info Team · Written with AI ·
Researchers submitted Agent-Reflex to arXiv on October 13, proposing a training method for language agents that encounter faulty tool responses or mid-task corrections from users. Their Qwen3-8B-based system scores 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, the paper reports.
The work targets a practical weakness in tool-using agents: completing a task when services behave as expected is different from recovering when a tool times out, returns misleading information or gives an error. The authors’ method trains recovery as a learned capability, rather than relying only on an agent to critique its own work while it is running.
“Further analyses show that Agent-Reflex improves agents’ robustness to noise while maintaining stable training dynamics, supporting scalable agent training.”
Guanting Dong and co-authors, the authors of the paper
From observed errors to reusable recovery skills
Agent-Reflex has three stages. Environment Reflection Skill Discovery, or RSD, builds executable environments around 557 high-use MCP service themes, introduces tool-return failures, and compares successful and failed agent runs. From those comparisons, the researchers extract reusable rules for handling recurring mistakes.
The source data includes 24,510 error returns among 177,700 tool returns collected by replaying agents across nine public benchmarks. The team groups errors into eight families and mines 1,170 reflection skills across the 557 environments. To produce contrasting trajectories, it samples nine language models from five model families, ranging from 4 billion to 284 billion parameters.
Three kinds of correction, then on-policy training
In the second stage, Reflective Supervised Fine-Tuning (ReSFT) turns the mined skills into examples of three behaviors: diagnosing and repairing a mistake, recovering from a flawed initial plan, and responding to a user’s correction. The authors say they keep only continuations that pass the environment’s task verifier. In the self-repair and recovery examples, the skill prompt is removed from the stored input, so the model must learn the corrective behavior rather than depend on seeing that private guidance at use time.
The third stage, On-Policy Reflective Policy Optimization (ORPO), combines reinforcement learning on clean and noisy task runs with reflection-guided training at the token level. The model’s decisions are scored both with and without an environment-specific skill, and the resulting signal gives more weight to tokens the skill supports. The aim is to teach the policy to choose corrective actions during its own runs, not merely to memorize successful repair examples.
Scores improve against other trained 8B systems
On the five long-horizon benchmarks, Agent-Reflex posts a 58.6 overall average, compared with 40.2 for the Qwen3-8B starting model and 50.9 for EnvScaler-8B. The benchmark set includes VitaBench, τ²-Bench, BFCL V4, ACEBench and Tool Sandbox. The comparison is not a claim that Agent-Reflex beats every model in the table: the paper reports higher overall results for Gemini 3 Pro, Claude-4.5-Sonnet and GPT-5.2.
On AgentNoiseBench, Agent-Reflex scores 29.1 overall, ahead of the paper’s 25.9 for NoisyAgent-8B and 18.2 for EnvScaler-8B. The benchmark tests both noisy tool responses and user interventions. The paper says AgentNoiseBench’s own noise taxonomy and injector were not used in training; its perturbations were designed independently, so the evaluation measures transfer to unseen perturbations.
A fixed skill library needs upkeep
The paper’s results come from benchmark experiments, not evidence of improved outcomes in deployed services. Its authors also identify a maintenance constraint: the skill library is mined offline from a fixed snapshot and does not automatically adjust when tool interfaces, service behavior or environment dynamics change. Keeping it aligned may require periodic re-mining or other maintenance.
Agent-Reflex’s performance therefore rests on both the learned policy and the continued relevance of its environment-specific skills. The authors provide code with the paper, while noting that changing tool behavior can make those mined recovery rules stale.