OPIUM: Mitigating Steering Externalities and Over-Refusal via Dual Objective Latent Optimization

TL;DR - OPIUM is a training-free method that sanitizes activation-steering vectors to reduce safety degradation and excessive refusal. It improves the safety–utility tradeoff by directly optimizing model activations against desired and protected reference behaviors.

• Uses representation matching across separate prompt sets to preserve intended steering effects while correcting failures.
• Addresses both utility vectors that weaken safety and refusal vectors that reject benign prompts.
• Optimizes a replacement steering vector without retraining the underlying model.
• Outperforms vanilla steering and directional ablation across the evaluated settings.