Audited Skill-Graph Self-Improvement for Agentic LLMs

This paper addresses critical security and governance challenges in self-improving agentic LLMs. It proposes a framework, ASG-SI, that focuses on creating auditable and verifiable improvements. The core idea is to treat self-improvement as a process of compiling an agent into a growing skill graph, ensuring that each improvement is extracted from successful trajectories, normalized into a skill with a clear interface, and validated through verifier-backed checks. This approach aims to mitigate issues like reward hacking and behavioral drift, making the self-improvement process more transparent and manageable. The integration of experience synthesis and continual memory control further enhances the framework's scalability and long-horizon performance.