RUC-NLPIR Releases Agent-Reflex Framework: Enhancing Agent Reflection and Robustness
Summary:RUC-NLPIR has introduced Agent-Reflex, a novel training framework designed to enhance the reflective capabilities of language agents. The framework operates in three stages: Environment Reflection Skill Discovery (RSD), Reflective Supervised Fine-Tuning (ReSFT), and On-Policy Reflective Policy Optimization (ORPO). Evaluated on multiple benchmarks, Agent-Reflex achieved an average score of 58.6 on long-horizon tool-use tasks and 29.1 on AgentNoiseBench, outperforming strong environment-scaling an
Background and Motivation
As large language models (LLMs) evolve from single-turn chatbots into interactive agents capable of long-horizon reasoning, the challenge of enabling reflective capabilities in these agents becomes critical. In real-world deployments, tool services may return noisy feedback, and users might intervene to correct the agent's behavior. However, existing methods provide limited supervision for reflection learning and failure recovery.
The Agent-Reflex Framework
Agent-Reflex is a novel training framework designed to incentivize proactive reflection in agents. It operates in three stages:
-
Environment Reflection Skill Discovery (RSD):
- Perturbations based on tool feedback are injected into the environment scaling to generate a diverse pool of LLM capabilities.
- Matched success-failure pairs are mined, and reusable reflection meta-skills are summarized.
-
Reflective Supervised Fine-Tuning (ReSFT):
- Each skill is converted into supervision for self-reflection, skill recovery, and user-centric correction.
- Complementary reflective behaviors are strengthened to enhance the overall reflective capabilities of the agent.
-
On-Policy Reflective Policy Optimization (ORPO):
- Combines agentic reinforcement learning with reflection-guided policy distillation to further elicit proactive reflection in noisy interactions.
Experiments and Results
Built on Qwen3-8B, Agent-Reflex achieved an average score of 58.6 on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, outperforming strong environment-scaling and reflection baselines. The results demonstrate the framework's effectiveness in the following areas:
- Noise Robustness: The agent can reflect and adjust strategies more effectively in noisy environments.
- Training Stability: Maintains stable training dynamics, supporting large-scale deployment.
Technical Mechanism Analysis
The core of Agent-Reflex lies in its multi-stage reflection mechanism:
- RSD Stage: Builds a rich reflection skill library through perturbation injection and skill mining, providing a solid foundation for subsequent fine-tuning.
- ReSFT Stage: Converts reflection skills into supervision signals, ensuring the agent learns effective reflective behaviors during training.
- ORPO Stage: Introduces reflection-guided policy distillation into reinforcement learning, further enhancing the agent's reflective capabilities.
Engineering Trade-offs and Performance
- Strengths:
- Improves the agent's robustness to noise.
- Maintains training stability, supporting large-scale deployment.
- Weaknesses:
- Increases the complexity of the training process, potentially requiring more computational resources.
- Sensitive to the choice of the initial model, requiring a high-quality pre-trained model.
Developer Implementation and Deployment Recommendations
- Model Selection: Use a high-quality pre-trained model as the foundation to maximize the benefits of Agent-Reflex.
- Training Resources: Due to the multi-stage reflection mechanism, training may require more computational resources. It is recommended to train on a GPU cluster.
- Application Scenarios: Suitable for long-running agent tasks such as dialogue systems, virtual assistants, and automated workflows.
Conclusion
The Agent-Reflex framework offers a new approach to enhancing the reflective capabilities and robustness of agents, with broad application prospects.
— END —Source: Hugging Face Daily Papers (2026-10-13)
Tags: #Agent-Reflex #Intelligent Agents #Reflection Learning #Reinforcement Learning #Noise Robustness
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments