EVA: General Alignment Against Jailbreaks via Model Editing

The EVA framework achieves jailbreak defense for LLMs and VLMs through direct model editing, reframing safety alignment as a precise knowledge correction task and performing targeted edits on specific neurons susceptible to harmful instructions. The method locates and modifies these neurons to neutralize harmful behavior, and the paper's experiments show that its effectiveness in mitigating jailbreaks on LLMs and VLMs is superior to baselines. It does not require retraining a large number of parameters, preserves general reasoning ability as much as possible, and is suitable for post-deployment safety alignment.