OPIUM: Mitigating Steering Externalities and Over-Refusal via Dual Objective Latent Optimization
OPIUM:通过双目标潜在优化减轻引导外部性和过度拒绝
AI总结 研究激活引导向量存在的外部性问题,提出无训练方法OPIUM,通过表示匹配净化引导向量,在给定参考行为下优化新向量,改善安全与效用权衡,减轻激活引导的有害副作用。
Comments Accepted to the Mechanistic Interpretability Workshop at the 43rd International Conference on Machine Learning (ICML 2026)