A Systematic Investigation of RL-Jailbreaking in LLMs
LLMs中RL越狱的系统性研究
机构 * University of California, Berkeley(加州大学伯克利分校)
AI总结 本文首次系统分解RL越狱框架,通过分析奖励函数、动作空间、回合长度等环境形式化因素和算法措施,发现密集奖励和延长回合长度是越狱成功的主要驱动因素,并提供了提升RL越狱效率及强化模型防御的工具。
Comments Warning: This paper may contain unfiltered and potentially offensive jailbreaking examples. Accepted at the Second Workshop on Agents in the Wild: Safety, Security, and Beyond (AIWILD) at ICML 2026