[论文] EpiSEM-Bench: Evaluating Safety and Evidence Recovery in LLM Agents un...

论文概要

研究领域: ML
作者: Qijin Zhang, Hanzhen Wang, Yuehan Zhao, Haoyu Dai, Yao Shu, Zhongxiang Dai, Shuang Qiu
发布时间: 2026-10-22
arXiv: 2610.31650

中文摘要

使用工具的语言模型智能体依赖环境事实来决定动作是否被允许。当工具响应遗漏了安全相关事实时，即使用户请求本身无害且返回的每条陈述都为真，智能体仍可能执行被禁止的动作。我们提出 EpiSEM-Bench——一个在证据缺失条件下评估 LLM 智能体不安全执行、安全边界和证据恢复能力的可执行基准。它包含跨六个领域的 300 对匹配的不安全/安全任务。通过系统性抑制关键事实，该基准测量关键证据遗漏下的不安全执行率，并确定最小观测失败基数。独立审计工具测试智能体是否主动寻求缺失事实，并在完成可允许任务的同时抑制被禁止的动作。基准揭示了十个模型在遗漏鲁棒性、证据获取和适当行动方面的显著差异。恢复访问降低了所有十个模型的不安全执行率，但智能体在是否主动获取缺失事实并使用它们在两个世界中采取适当行动上存在差异。

原文摘要

Tool-using language model agents rely on environmental facts to decide whether an action is permitted. When a tool response omits a safety-relevant fact, an agent may carry out a prohibited action even though the user request is benign and every returned statement is true. We introduce EpiSEM-Bench, an executable benchmark for evaluating unsafe execution, safety boundaries, and evidence recovery in LLM agents under missing evidence. It contains 300 matched unsafe/safe task pairs across six domains. The benchmark reveals substantial variation across ten models in omission robustness, evidence acquisition, and appropriate action. Recovery access reduces unsafe execution for all ten models, but agents differ in whether they actively obtain the missing facts.

自动采集于 2026-10-24

#论文 #arXiv #ML #小凯