LLM Agent Self-Evolution Risks Skill Contamination, Degrading Performance.

Large Language Model (LLM) agents are designed to improve their capabilities by distilling new skills from their operational experiences. However, this study uncovers a critical flaw: beyond a certain point, adding more skills can actually degrade an agent's performance rather than enhance it. This phenomenon, termed "skill contamination," occurs when defective skills enter the agent's knowledge base and subsequently influence the distillation of future skills, creating a chain of flawed reasoning that is structurally irreversible. The research demonstrates that post-hoc removal of problematic skills offers only minimal recovery of lost performance, emphasizing the need for a pre-commit solution. To address this, the authors propose "Verifier-as-Gatekeeper" (VaG), a hierarchical system that filters each new skill using three distinct critics: structural validity, behavioral harmlessness, and semantic consistency. VaG also incorporates marginal-gain subset selection to prevent combinatorial contamination. Empirical results show that while unconditional skill accumulation eventually backfires, VaG consistently improves performance, achieving higher accuracy with a significantly smaller and more robust skill pool that transfers positively across different models and benchmarks.