relearning-attacks

This paper introduces Minor Component Unlearning (MCU), a novel approach to LLM unlearning that targets minor components in representations to resist relearning attacks. It addresses the vulnerability of existing methods by focusing on robust directions within the model's spectral structure.