State-Dependent Safety Failures in Multi-Turn Language Model Interaction
多轮语言模型交互中的状态依赖性安全故障
机构 * School of Cyber Science and Technology, University of Science and Technology of China, China(中国科学技术大学信息科学与技术学院) ; Nanyang Technological University, Singapore(南洋理工大学) ; Tsinghua University, Beijing, China(清华大学) ; Beijing Electronic Science and Technology Institute, Beijing, China(北京电子科技研究所)
专题命中 越狱攻击 :safety(title,abstract);alignment(abstract);jailbreak(abstract_cn);分类 cs.AI
AI总结 本文从状态空间视角研究多轮交互中的安全故障,提出STAR框架分析对话历史作为状态转移操作,揭示对齐模型在多轮交互中安全边界穿越的机制。
Comments 9 pages, accepted at the International Conference on Machine Learning (ICML) 2026