Verifier-as-Gatekeeper: Preventing Skill Contamination in Self-Evolving LLM Agents

A recent paper on arXiv (2608.05810) discusses self-evolving large language model (LLM) agents that enhance their abilities by extracting reusable skills from their execution paths. It highlights a non-monotonic evolution of capabilities: when the pool of skills surpasses a critical size, performance can decline with the addition of new skills. The authors introduce the term 'capability-contamination phase transition' to explain this phenomenon, attributing it to a structural issue where defective skills, once included in decision-making, serve as reference points for future skill development, resulting in contamination across iterations. They show that this contamination is structurally irreversible, as eliminating a flawed skill does not remove the inherited poor reasoning in subsequent skills. This leads to the advocacy for a pre-commitment strategy for skill acceptance, proposing the Verifier-as-Gatekeeper (VaG) model, which consists of three diverse critics that evaluate structural validity prior to skill admission. The paper can be found on arXiv under the identifier 2608.05810.