arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.00061cs.CYcs.AI
收藏

简单的角色分配在安全对齐中表现尤为有效

Simple Role Assignment is Extraordinarily Effective for Safety Alignment

University of California, Los Angeles(加州大学洛杉矶分校);Tianjin University(天津大学);Peking University(北京大学);Zhejiang University(浙江大学);Beijing Institute of General Artificial Intelligence(北京一般人工智能研究院)

Zhou Ziheng, Jiakun Ding, Zhaowei Zhang, Ruosen Gao, Yingnian Wu, Demetri Terzopoulos, Yipeng Kang, Fangwei Zhong, Junqi Wang

AI总结:

本研究提出基于角色条件化的安全对齐方法,通过角色隐含编码价值观和认知模式,有效降低不安全输出,适用于多种AI对齐任务。

摘要:

基于原则的对齐方法常常缺乏上下文敏感性和完整性。基于心灵理论,我们提出角色条件化作为一种简洁的替代方案:社会角色(例如母亲、法官)隐含地编码了价值观以及应用这些价值观所需的认知模式。我们引入了一种无需训练的流程,包括角色条件生成器和迭代基于角色的批评者以进行优化。在五个模型家族中,我们的方法在多个基准测试中均优于基于原则、思维链(CoT)和其他基线方法。值得注意的是,它在WildJailbreak基准测试中将不安全输出率从81.4%降至3.6%(使用DeepSeek-V3)。不仅适用于常见的安全基准,它还适用于代理安全任务。这些结果确立了角色分配作为AI对齐和LLM-as-a-Judge构建的强大、可解释的范式。

英文摘要:

Principle-based alignment often lacks context sensitivity and completeness. Grounded in Theory of Mind, we propose role conditioning as a compact alternative: social roles (e.g., mother, judge) implicitly encode both values and the cognitive schemas required to apply them. We introduce a training-free pipeline featuring a role-conditioned generator and iterative role-based critics for refinement. Across five model families, our approach consistently outperforms principle-based, Chain-of-Thought (CoT) and other baselines across benchmarks. Notably, it reduces unsafe outputs on the WildJailbreak benchmark from 81.4\% to 3.6\% with DeepSeek-V3. Not only for common safety benchmarks, it consistently applies for agentic safety tasks. These results establish role assignment as a powerful, interpretable paradigm for AI alignment and LLM-as-a-Judge construction.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑