Tag
This paper investigates how a biased LLM judge silently disables skill retirement in self-evolving agents, showing that false-pass bias across a sharp threshold prevents contribution-based retirement and that the failure is universal across domains, detectable only through a defect-injection audit.
This paper identifies 'library drift' as a silent failure mode in self-evolving LLM skill libraries, where unbounded skill accumulation causes retrieval degradation and performance stagnation. It provides trace-level diagnostics and a verified governance recipe that lifts pass@1 from 0.258 to 0.584 on MBPP+ hard-100.