OPIUM: Mitigating the side effects and over-refusal of activation steering through dual-objective latent optimization

OPIUM is a training-free method that purifies large language models' steering vectors through dual-objective latent optimization, mitigating the side effects and over-refusal of activation steering. While preserving the downstream representations of the targeted intervention, the method makes prompts that the original vector fails on match safer reference behavior, avoiding utility vectors weakening safe behavior and refusal vectors causing over-refusal on benign prompts. In the steering side-effect and over-refusal settings, its safety-utility trade-off outperforms ordinary steering and direction ablation.