Nullify preprint reports selective LLM forgetting without weight updates

Researchers report in a preprint that Nullify, a method designed to suppress specific memorized information in large language models, matches or surpasses established forgetting baselines on TOFU and MUSE while preserving model utility with near-lossless results. Nullify intervenes when the model generates answers, using steering vectors to redirect privacy-related internal activations away from memorized responses. A mathematical constraint keeps activations for retained queries essentially unaffected. The method avoids training and changes to model weights entirely. Its intervention operates during inference, rather than erasing information through updates to the model's parameters.