When Self-Evolution Backfires: Pre-Commit Gating against Skill Contamination in LLM Agents

The authors study self-evolving LLM agents that accumulate reusable skills from execution trajectories. They find that this accumulation is not monotonic: past a critical pool size, adding new skills hurts performance instead of helping. They formalize this as a capability-contamination phase transition, tracing its structural cause to cross-round contamination chains—once a defective skill enters the decision context, it becomes reference material for distilling later skills. They show this contamination is structurally irreversible: post-hoc removal of a source skill cannot erase flawed reasoning inherited by descendants, and rollback recovers only a small fraction of lost performance. This motivates skill admission as a pre-commit necessity and leads to Verifier-as-Gatekeeper (VaG), a progressive trust hierarchy with three heterogeneous critics—structural validity, behavioral harmlessness, and semantic consistency—filtering each skill individually, plus marginal-gain subset selection to remove combinatorial contamination before skills reach runtime. On Terminal-Bench 2, unconditional accumulation peaks then degrades, giving back most gains as the pool grows, while post-hoc removal recovers only a small drop. VaG improves every round, reaching 72% pass@1 with a roughly 5x smaller pool, and its frozen skill pool transfers positively to four other backbones and a second benchmark without re-evolution. Ablations confirm the three critics are complementary and non-substitutable, intercepting largely disjoint classes of harmful skills.