The Blind Curator: How a Biased Judge Silently Disables Skill Retirement in Self-Evolving Agents

2026-07-08

AI Summary

This study addresses a critical failure mode in reference-free agent training: biased reward judges based on large language models (LLMs) exhibit a systematic “false-pass” bias that completely disables skill elimination mechanisms, leading to performance degradation in the agent’s skill repertoire. The work demonstrates for the first time that this bias is not mere stochastic noise but a structured flaw that universally disables淘汰 across domains and remains undetectable in aggregate metrics. Through corrupted reward modeling, causal interventions, defect-injection audits, and multi-task validation, the authors show that symmetric noise does not impair elimination, whereas false-pass errors beyond a critical threshold cause total failure. They further introduce a low-cost auditing method capable of providing early warnings when reward judges enter hazardous operational regimes.

Abstract

A self-evolving agent retires its bad skills by watching them fail, so what happens when the judge cannot see the failures? Skill retirement is the structural constraint that keeps a growing library from drifting below the no-skill baseline, but its guarantee assumes an unbiased reward, which is false for the LLM judges that reference-free tasks force upon us. We show that a biased judge does not merely add noise; it \emph{silently switches off the curator}. We make this precise with a corrupted-reward analysis and, isolating the causal channel by injecting corruption on top of a deterministic reward, a behavioral study on a reference-free report-writing testbed with a code-generation cross-check. Symmetric noise leaves retirement intact, but \emph{false-pass} bias (failures slipping through as passes) disables contribution-based retirement past a sharp threshold that no amount of data can cross. Separating genuine retirement from cap-eviction churn shows this \emph{mechanism} failure is universal, holding across domains and failure rates and sparing only near-zero-false-pass, verifier-like graders. The downstream \emph{outcome}, though, is regime-dependent: eval quality degrades only where the same corruption also starves skill synthesis, and otherwise holds steady, so the disabled curator is \emph{silent}, surfacing in no aggregate metric. The contribution is a behavioral safety result, not a performance one. A cheap defect-injection audit then tells an operator, before deployment, which side of the threshold their judge occupies.

Problem

• skill retirement
• biased judge
• self-evolving agents
• false-pass bias
• reference-free evaluation

Innovation

• skill retirement
• biased reward
• false-pass bias
• self-evolving agents
• behavioral safety