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

2026-08-06

AI Summary

This work identifies and formalizes a previously uncharacterized phenomenon—“skill contamination phase transition”—in self-evolving large language model agents, wherein performance degrades once the skill pool expands beyond a critical scale due to defect-laden skills inducing cross-iteration contamination that is structurally irreversible. To mitigate this, the authors propose a pre-commit gating mechanism that employs a triadic heterogeneous validator framework—assessing structural validity, behavioral harmlessness, and semantic consistency—to filter candidate skills prior to integration. Coupled with a marginal-gain-driven subset selection strategy, this approach effectively halts contamination accumulation. Evaluated on Terminal-Bench 2, the method achieves a 72% pass@1 rate while compressing the skill pool by 5×; notably, the frozen skill pool demonstrates zero-shot transferability across four backbone architectures and an additional benchmark.

Abstract

Self-evolving agents accumulate capability by distilling reusable skills from their execution trajectories, but we find this process is not monotonic: past a critical pool size, newly added skills degrade performance instead of improving it. We formalize this capability-contamination phase transition and trace it to a structural cause: once a defective skill enters the decision context, it becomes reference material for distilling later skills, forming cross-round contamination chains. We further show the contamination is structurally irreversible: removing a source skill after the fact cannot erase the flawed reasoning its descendants have already inherited, so post-hoc rollback recovers only a small fraction of the lost performance. This makes skill admission a pre-commit necessity rather than a post-hoc fix, and motivates Verifier-as-Gatekeeper (VaG): a progressive trust hierarchy whose three heterogeneous critics - structural validity, behavioral harmlessness, and semantic consistency - filter each skill individually, coupled with a marginal-gain subset selection that removes combinatorial contamination at the top tier before skills reach the runtime context. On Terminal-Bench 2, unconditional accumulation rises to a peak and then degrades, giving back most of its gains as the pool keeps growing, and post-hoc removal of the culprit skills recovers only a small part of the drop - the empirical signature of irreversibility. In contrast, VaG improves every round, reaching 72% pass@1 with a pool roughly 5x smaller, 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 mutually non-substitutable, each intercepting a largely disjoint class of harmful skills.

Problem

• skill contamination
• self-evolving agents
• capability degradation
• irreversible contamination
• LLM agents

Innovation

• skill contamination
• self-evolving agents
• pre-commit gating
• Verifier-as-Gatekeeper
• irreversible degradation