检验工具调用型大模型代理的评估可靠性，发现自动评分不可靠，错误会传播，但可实时拦截幻觉。
Auditing Automated Evaluation, Error Propagation, and Runtime Mitigation in Tool-Using Language Agents
- 用1.47万条执行轨迹构建诊断基准，对比自动评分与人工标注差异
- 自动评分仅随机准确（kappa=0.049），GPT-4o-mini表现最佳（kappa=0.567）
- 轻量级运行时拦截器可降低幻觉24个百分点，适合开源模型优化
工具调用型大语言模型代理的自动化评估普遍被认为可靠，但极少与人工标注验证。我们提出AgentProp-Bench，包含14,750条来自十三个大模型代理（九个专有、四个开源）在四个领域的执行轨迹，用于审计三个问题：第一，基于子串启发式的自动评分与人工标注仅达随机水平（Cohen's kappa = 0.049），而三模型集成达中等一致（kappa = 0.432），单个GPT-4o-mini判别器最强（kappa = 0.567）；双人工标注者一致性几乎完美（kappa = 0.835）。第二，在经验证的评判下，参数级错误导致最终答案错误的概率约为0.62，且在专有与开源模型间一致，模型拒收损坏输入与恢复能力统计独立。第三，多个代理虚构工具调用——声称获取了未实际获得的结果（高达24%的幻觉不可被端到端分数察觉），轻量级运行时拦截器可在可调操作点上将幻觉降低24个百分点，对开源模型净收益为正。所有代码、数据、轨迹及人工标签已公开于https://github.com/bhaskargurram-ai/agenthallu-bench。
原文摘要 · Abstract (English)
Automated evaluation of tool-using large language model (LLM) agents is widely assumed to be reliable, yet this assumption is rarely validated against human annotation. We present AgentProp-Bench, a diagnostic benchmark of 14,750 execution traces from thirteen LLM agents (nine proprietary, four open-weight) across four domains, and use it to audit three questions. First, substring-heuristic judging of agent outputs agrees with human annotation only at chance level (Cohen's kappa = 0.049 against each of two annotators), while a three-LLM ensemble reaches moderate agreement (kappa = 0.432) and a single GPT-4o-mini judge is in fact the strongest (kappa = 0.567); dual-annotator agreement is almost perfect (kappa = 0.835). Second, under validated judging a parameter-level error propagates to a wrong final answer with human-calibrated probability approximately 0.62, replicated across proprietary and open-weight models, and a model's ability to reject corrupted inputs and to recover from them are statistically indepen1). Third, severalagents fabricate tool executions -- asserting tool-derived renever obtained (up toilure invisible toend-to-end scores, and a lightweight runtime interceptor reduhallucination on everto 24 percentagepoints) at a tunable operating point that is net-positive on open-weight models. Ini-2.0-Flash, whichrarely emits an injectable tool call. All code, data, traces, and human labels are relehttps://github.com/bhaskargurram-ai/agenthallu-bench
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。