Nullify proposes steering language models away from private answers without retraining

Nullify aims to stop language models from reproducing selected private information by intervening as they answer, without changing their weights. Wei Zhai and eight co-authors presented the method in a preprint submitted to arXiv on 7 October 2026. It is a proposal for model unlearning—the effort to suppress specific knowledge while preserving a model’s usefulness elsewhere.

The researchers report that Nullify matches or beats established comparison methods on forget quality across two benchmarks, TOFU and MUSE, with what they call “near-lossless preservation of model utility.” Those are the authors’ results, and the abstract supplies neither numerical scores nor the names of the competing methods. It offers a promising headline result, but no way to judge the size of the reported advantage.

The intervention happens during inference, when the model generates an answer. Steering vectors redirect internal activations associated with private information away from memorized answers. A mathematical restriction—the null-space constraint in the paper’s title—is intended to leave activations for queries outside the forgetting target essentially unchanged. The authors describe this as a training-free alternative to methods that fine-tune model parameters.

The distinction is important: Nullify changes the answering process while leaving the weights untouched. That is not, by itself, evidence that stored information has been permanently erased. The abstract also does not describe attempts to recover the targeted information, leaving that question open in the available evidence.