Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries
The paper discusses a phenomenon called 'library drift' in self-evolving LLM skill libraries, which leads to performance issues. It identifies the causes of this drift and proposes a governance framework to mitigate its effects. The authors present empirical evidence showing significant improvements in performance when their recommendations are implemented.
- ▪Library drift results in unbounded skill accumulation that degrades retrieval performance.
- ▪LLM-authored skills show no performance gain compared to a 16.2pp gain from human-curated skills.
- ▪The proposed governance framework includes outcome-driven retirement and bounded active capacity, leading to improved performance metrics.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.19576 (cs) [Submitted on 19 May 2026] Title:Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries Authors:Xing Zhang, Yanwei Cui, Guanghui Wang, Ziyuan Li, Wei Qiu, Bing Zhu, Peiyang He View a PDF of the paper titled Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries, by Xing Zhang and 6 other authors View PDF HTML (experimental) Abstract:Self-evolving skill libraries face a silent failure mode we term \emph{library drift}: unbounded skill accumulation without outcome-driven lifecycle management causes retrieval degradation, false-positive injections, and performance stagnation.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
