Nullify: Null-Space Activation Steering for Training-Free LLM Unlearning
Nullify is a training-free method for unlearning information from LLMs by steering activations during inference. It uses linear corrections applied to the residual stream to push privacy-related activations away from memorized answers. Crucially, it maintains model utility by constraining these corrections to the null space of retained knowledge, ensuring kept queries remain unaffected.