knowledge-removal

The paper argues that unlearning in LLMs should be goal-dependent, proposing a cosine-based meta-learned variant of RMU for dangerous knowledge and a multi-layer objective with probe directions for toxicity, achieving strong results across four 7-8B models.