[论文] What the Trace Shows and What the Label Says: Auditing Behavioral Dete...

论文概要

研究领域: ML
作者: Idunnuoluwa A. Adeniji, Shilpika, Michael E. Papka
发布时间: 2026-10-24
arXiv: 2610.30491

中文摘要

LLM 智能体评估常把执行 trace 归约为标签：最终答案正确性、工具验证、故障传播。这些标签是解析器、行为检测器与 harness 规则产出的测量，未必完整捕捉 trace 所显示的内容。本文在基于 GSM8K 的工具使用场景中研究这一问题：智能体调用计算器，其输出偶发损坏——工具实际返回了什么、模型后来用了什么、检测器报告了什么，都有清晰的 ground truth。审计三个 trace 级检测器 × 六个生产模型 × 每模型 30 题：最终正确性、可见的再验证与故障传播会以系统性方式互相矛盾。一个关键失效模式是静默工具输出覆盖——计算器返回损坏值，但模型稍后未再调用工具却使用了另一个（通常正确的）值：坏值停止传播，而验证检测器记录不到任何重查。对 90 条 trace 的人工复核与自动标签在最终正确性（67/67）、故障影响（19/19）、故障相关性（18/18）上完全一致，验证项仅 67/81，14 处分歧归入四个重复模式。执行预算也会改变报告结果：两个模型的消融中，一者的正确数从 21/30 到 27/30 随步数与 token 预算漂移，另一者稳定在 27/30–29/30。这些发现主张测量透明：自动标签应附带其 trace 证据、检测器问题是否适用于该 trace、运行是否被执行上限截断、以及任何人工修正。配套的 trace 查看器让这些信息可与报告结果一并检视。

原文摘要

LLM agents are often evaluated by reducing execution traces to labels such as final-answer correctness, tool verification, and fault propagation. These labels are measurements produced by parsers, behavioral detectors, and harness rules, and they may not fully capture what the trace shows. We study this in a tool-use setting built on GSM8K, where agents call a calculator whose outputs are occasionally corrupted, giving clear ground truth for what the tool returned, what the model later used, and what the detectors reported. We audit three trace-level detectors across six production models and 30 problems per model. Final correctness, visible re-verification, and fault propagation can disagree in systematic ways. A key failure mode is silent tool-output override: the calculator returns a corrupted value, but the model later uses a different, often correct, value without calling the tool again, so the bad value stops propagating while a verification detector records no re-check. Human review of 90 traces agrees perfectly with the automated labels for final correctness (67/67), fault impact (19/19), and fault relevance (18/18), but only 67/81 for verification; all 14 disagreements fall into four recurring patterns. Execution limits also change the reported result: in an ablation across two models, one ranges from 21/30 to 27/30 correct as step and token budgets change, while the other stays between 27/30 and 29/30. These findings motivate measurement transparency: an automated label should come with the trace evidence behind it, whether the question the detector asks applied to that trace, whether the run was cut off by an execution limit, and any human corrections. Our trace viewer makes this information inspectable alongside the reported results.

自动采集于 2026-10-24

#论文 #arXiv #ML #小凯