Self-Healing Framework for Reliable LLM-Based Autonomous Agents
A study available on arXiv (2605.06737) presents a self-healing framework focused on reliability for autonomous agents powered by LLMs. This framework tackles unpredictable issues, including hallucinations, execution mistakes, and inconsistent reasoning. It combines failure detection, reliability evaluation, and automated recovery processes. The authors categorize different types of failures and propose a quantitative model for assessing reliability. An abnormal behavior detection method monitors execution patterns and output consistency. Additionally, the self-healing mechanism employs adaptive replanning and corrective prompting to recover from failures dynamically. The framework has been tested within a multi-agent workflow environment.
Key facts
- arXiv paper 2605.06737 proposes a self-healing framework for LLM-based autonomous agents.
- Framework addresses failures including hallucinations, execution errors, and inconsistent reasoning.
- Integrates failure detection, reliability assessment, and automated recovery.
- Defines a taxonomy of failure types and a quantitative reliability assessment model.
- Failure detection method uses execution patterns and output consistency.
- Self-healing mechanism uses adaptive replanning and corrective prompting.
- Implemented and evaluated in a multi-agent workflow environment.
- Aims to improve reliability of LLM-based agents in complex software systems.
Entities
Institutions
- arXiv