Nullify preprint proposes inference-time steering to make language models forget

A preprint posted to arXiv on 7 October 2026 describes Nullify, a way to make a large language model stop reproducing selected information without retraining it. The nine authors, led by Wei Zhai, call it a training-free, non-destructive activation steering method for LLM unlearning.

Unlearning aims to erase specific knowledge, such as private details absorbed during pre-training, so that a model no longer leaks it while staying useful for everything else. The authors say existing methods struggle to balance how well they forget against how much utility they keep. They also say these methods typically cost a lot of computation because they fine-tune parameters.

What is new is where the intervention happens. According to the abstract, Nullify applies steering vectors while the model is answering, redirecting privacy-related internal activations away from the answers the model has memorized. A null-space constraint is meant to leave activations for queries outside the forget target essentially unchanged. The weights are never updated, so the approach is described as a plug-and-play intervention rather than a retraining step.

The performance claims are the authors' own. They report that on two unlearning benchmarks, TOFU and MUSE, Nullify matches or surpasses established baselines in forget quality, with near-lossless preservation of model utility. The abstract gives no scores, no list of baselines and no margins, so the size of any advantage cannot be judged from it.

Much remains unconfirmed. The only source available to us is the arXiv abstract page, and we found no independent second source on this paper. The page does not say the work has been peer reviewed, and we have not read the full text, code or experimental details. Whether the method holds up on other models, other kinds of private information, or under attempts to extract the supposedly forgotten material is not addressed in the abstract.

It matters because the usual route to unlearning means updating a model, which is expensive and can degrade it. An inference-time approach that leaves weights untouched would be cheaper to apply and easier to roll back, if the reported results survive outside the authors' evaluation. For now it is a proposal with self-reported results, not an established technique.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.