Agentic AI
Library Drift: A Silent Failure Mode in Self-Evolving LLM Skill Systems
Researchers identify and fix "library drift," where unbounded skill accumulation in self-evolving LLM systems degrades performance without warning.
A new arXiv paper identifies a critical failure mode in self-evolving LLM skill libraries called "library drift"—a phenomenon where unchecked skill accumulation causes retrieval degradation, false-positive injections, and performance stagnation, all while remaining invisible to traditional performance metrics.
The Problem#
The research reveals that LLM-authored skills provide negligible gains (+0.0 percentage points) on SkillsBench while human-curated skills deliver +16.2pp, yet the root cause of this degradation remained unclear. Library drift operates as a silent failure: the system appears functional at the end-task level even as internal mechanisms deteriorate.
Diagnosis and Solution#
The authors provide three key contributions:
-
Reproducible Trigger: Ablation studies isolate drift's mechanism, showing that disabling skill injection yields a floor (+0.002 gain) while premature retirement causes active harm (−0.019).
-
Trace-Level Diagnostics: An append-only evidence log tracks per-skill contribution scores, attribution verdicts, and router engagement metrics, making failures visible before they degrade end-task performance.
-
Verified Fix: A minimal governance recipe combining outcome-driven skill retirement, bounded active-cap constraints, and meta-skill authoring priors improves performance on MBPP+ hard-100 from a 0.258 baseline to a rolling mean of 0.584 over 100 evaluation rounds—a +0.328 gain.
The paper includes eight ablations decomposing which governance mechanisms are essential, offering a concrete playbook for diagnosing and preventing library drift in any self-evolving agent system.
