LLM评估指令遵守性时，结论稳定但推理不稳，易隐藏风险。
The Stability Trap: Evaluating the Reliability of LLM-Based Instruction Adherence Auditing
- 按指令类型分解，区分客观与主观任务，定位评估不稳定根源。
- 评判一致率超99%，但推理一致性最低仅19%，差异显著。
- 建议代码处理确定性逻辑，仅让LLM负责复杂语义判断。
企业在监管领域（如人力资源）应用生成式AI需可扩展且可复现的审计机制。尽管大语言模型（LLM）作为裁判具有可扩展性，但其对不同类型系统指令的合规性评估可靠性尚未验证。本研究探究：待测应用（AUT）的指令类型如何影响裁判评估的稳定性？我们提出受限指令分解框架，将指令分为客观与主观两类，分离导致评估不稳定的因素。在两个代表性人力资源类生成式AI应用中，评估了四种裁判架构在多次运行中的稳定性。结果揭示存在‘稳定性陷阱’：判据稳定性与推理稳定性严重偏离。尽管对客观和主观指令的判决一致性均超过99%，但对应解释轨迹差异极大。涉及定量分析（如字数统计）的客观指令，推理稳定性低至约19%，源于数值解释波动；主观指令的推理稳定性则在35%至83%间波动，受证据粒度影响，特定特征检查无法重现一致理由。而聚焦离散实体提取的客观指令则实现>90%的推理稳定性。研究证明，高判决一致性可能掩盖脆弱的推理过程。因此建议审计者严格限定自动化评估范围：将所有可确定性验证的逻辑交由代码执行，仅保留LLM裁判用于复杂语义评估。
原文摘要 · Abstract (English)
The enterprise governance of Generative AI (GenAI) in regulated sectors, such as Human Resources (HR), demands scalable yet reproducible auditing mechanisms. While Large Language Model (LLM)-as-a-Judge approaches offer scalability, their reliability in evaluating adherence of different types of system instructions remains unverified. This study asks: To what extent does the instruction type of an Application Under Test (AUT) influence the stability of judge evaluations? To address this, we introduce the Scoped Instruction Decomposition Framework to classify AUT instructions into Objective and Subjective types, isolating the factors that drive judge instability. We applied this framework to two representative HR GenAI applications, evaluating the stability of four judge architectures over variable runs. Our results reveal a ``Stability Trap'' characterized by a divergence between Verdict Stability and Reasoning Stability. While judges achieved near-perfect verdict agreement ($>99\%$) for both objective and subjective evaluations, their accompanying justification traces diverged significantly. Objective instructions requiring quantitative analysis, such as word counting, exhibited reasoning stability as low as $\approx19\%$, driven by variances in numeric justifications. Similarly, reasoning stability for subjective instructions varied widely ($35\%$--$83\%$) based on evidence granularity, with feature-specific checks failing to reproduce consistent rationale. Conversely, objective instructions focusing on discrete entity extraction achieved high reasoning stability ($>90\%$). These findings demonstrate that high verdict stability can mask fragile reasoning. Thus, we suggest that auditors scope automated evaluation protocols strictly: delegate all deterministically verifiable logic to code, while reserving LLM judges for complex semantic evaluation.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。