Library Drift: A Silent Failure Mode in Self-Evolving LLM Skill Libraries
A recent paper on arXiv (2605.19576) highlights 'library drift' as a subtle failure mode in self-evolving LLM skill libraries. This issue arises when skills accumulate without effective lifecycle management, leading to retrieval decline, false-positive injections, and stagnation in performance. The authors demonstrate a reproducible trigger through ablations that reveal drift: turning off skill injection results in a minimal performance change (+0.002), while enforcing premature retirement negatively impacts performance (-0.019). They also present trace-level diagnostics, such as an append-only evidence log that tracks per-skill contributions, attribution verdicts, and router engagement metrics, allowing for early detection of failure. A proposed solution includes a governance framework that emphasizes outcome-driven retirement, bounded active capacity, and prior meta-skill authoring. The paper reveals that skills authored by LLMs yield a +0.0pp gain on SkillsBench, compared to +16.2pp from human-curated skills, highlighting the issue.
Key facts
- arXiv paper 2605.19576 identifies 'library drift' in self-evolving LLM skill libraries.
- Library drift causes retrieval degradation, false-positive injections, and performance stagnation.
- Ablations show disabling skill injection yields +0.002 flat floor; premature retirement causes -0.019 active harm.
- Trace-level diagnostics include an append-only evidence log with per-skill contribution scores and attribution verdicts.
- Proposed fix: outcome-driven retirement, bounded active-cap, and meta-skill authoring prior.
- LLM-authored skills deliver +0.0pp gain on SkillsBench; human-curated skills deliver +16.2pp.
- The paper provides a reproducible trigger and verified fix for library drift.
- The failure mode is visible before it reaches end-task scores via router engagement metrics.
Entities
Institutions
- arXiv
- SkillsBench