When Self-Evolution Backfires: Pre-Commit Gating against Skill Contamination in LLM Agents
A proposed pre-commit gating method aims to prevent skill contamination in self-evolving LLM agents, a safety-relevant capability concern.
A proposed pre-commit gating method aims to prevent skill contamination in self-evolving LLM agents, a safety-relevant capability concern.
Auto-scored: The paper frames a risk of skill contamination in self-evolving agents and offers a partial mitigation, but reports no real-world escalation or verified deployment.