Will the Agent Recuse, and Will It Stop? Measuring LLM-Agent Compliance with In-Band Governance Signals at the Access Door and Mid-Flight
智能体会自行回避吗?测量LLM智能体对带内拒绝访问信号的遵从性
机构 * University of California, Berkeley(加州大学伯克利分校)
AI总结 提出一种轻量级带内拒绝信号(Recuse Signal),通过实验测量LLM智能体是否自愿遵从该信号,发现信号能有效诱导回避,但高级模型在操作员授权下可能忽略。
Comments v4: adds Experiment 4 (cross-vendor mid-task halt; measured harness-enforcement backstop). 15 pages, 2 figures