Agent-Reflex is a new training framework that teaches language agents to proactively notice and recover from noisy tool feedback and mid-task user corrections, rather than blindly pushing forward after something goes wrong. Built by researchers at the Gaoling School of Artificial Intelligence, Renmin University of China, and trained on Qwen3-8B, it posts 58.6 overall across five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, outperforming strong environment-scaling and reflection baselines. The code, models, and data are publicly released.
What problem is Agent-Reflex trying to solve?
Today’s agent training assumes a clean, stable environment, but real deployments are messy: tool services time out or error, and users interrupt to correct the agent’s course. The paper identifies two gaps in existing methods — a “lack of noisy feedback” in training environments and a “lack of user-guided correction,” since user simulators typically rely on standardized templates that poorly cover corrective instructions and diverse interaction styles. As a result, agents trained in idealized conditions are underprepared for failure detection and recovery when deployed in the real world.
How does Agent-Reflex actually work?
Agent-Reflex treats reflection as a learnable, transferable meta-skill rather than something agents pick up incidentally, and it does this through three stages. First, Environment Reflection Skill Discovery (RSD) anchors 557 high-usage real-world environment themes, builds database-backed executable environments and tasks, injects tool-return failures to simulate noise, samples trajectories from a capability-diverse LLM pool, and mines matched success–failure pairs into reusable reflection meta-skills. Second, Reflective Supervised Fine-Tuning (ReSFT) converts each mined skill into supervision targeting self-reflection, skill recovery, and user-centric correction. Third, On-Policy Reflective Policy Optimization (ORPO) combines agentic reinforcement learning under noisy interactions with reflection-guided on-policy distillation to further push proactive reflective behavior.
How does it compare to existing approaches?
Agent-Reflex is positioned against two prior families of environment-scaling methods: implicit world-model approaches like Simia and GenEnv, which generate feedback from LLMs but risk hallucination, and explicit approaches that synthesize executable tools and databases from real-world service themes. Agent-Reflex builds on the explicit, database-backed style but adds injected noise and mined reflection skills that neither prior group specifically targets.
| Approach | Feedback source | Noise handling | User correction |
|---|---|---|---|
| Simia / GenEnv (implicit) | LLM-generated from tool specs | Not addressed; risk of hallucination | Limited |
| Explicit executable-env methods | Real databases/executable tools | Not addressed | Limited |
| Agent-Reflex | Database-backed executable envs | Injected tool-return failures + mined meta-skills | User-centric correction via ReSFT |
What did the experiments show?
Agent-Reflex, built on Qwen3-8B, reached 58.6 overall across five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, and the authors report it surpasses strong environment-scaling and reflection baselines on six challenging agent benchmark suites. Further analysis found that Agent-Reflex improves agents’ robustness to noise while keeping training dynamics stable, which the authors say supports scalable agent training going forward.
Why does this matter for developers building agents?
For anyone deploying tool-using LLM agents in production, unreliable APIs and mid-task user corrections are everyday realities, not edge cases — this paper directly targets that gap with a reproducible training recipe rather than just a benchmark. Because the code, models, and data collection are openly released, developers can inspect or adapt the three-stage pipeline (RSD, ReSFT, ORPO) for their own agent-training setups instead of relying on idealized, noise-free simulators.