Agent-Reflex: Teaching Language Agents to Act Reflectively
Agent-Reflex trains tool-using agents to recover from noisy feedback and user corrections by practicing reflection directly.
The paper proposes a three-stage framework: discover reusable reflection skills from success-failure pairs, turn those skills into supervised training data, then refine the policy on noisy agent interactions. Built on Qwen3-8B, it reports 58.6 overall across five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench. The authors say the method beats environment-scaling and reflection baselines, while improving robustness without destabilizing training. Source: HF Daily Papers' note.
The paper proposes a three-stage framework: discover reusable reflection skills from success-failure pairs, turn those skills into supervised training data, then refine the policy on noisy agent interactions. Built on Qwen3-8B, it reports 58.6 overall across five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench. The authors say the method beats environment-scaling and reflection baselines, while improving robustness without destabilizing training. Source: HF Daily Papers' note.
score 4