Natively Unlearnable Large Language Models

A research team has introduced a new large language model architecture called NULLs, designed to let operators surgically remove the influence of specific training data sources without retraining the entire model. The work, posted to arXiv on June 11, 2026, addresses a persistent obstacle in machine unlearning: the contributions of different data sources are deeply entangled inside a model’s parameters, making clean removal difficult [1]. The proposed model class, short for Natively Unlearnable LLMs, trains a set of shared backbone neurons alongside a pool of sparsely activated “sinks” [2]. During training, information specific to a single source concentrates in its assigned sinks, while knowledge shared across sources accumulates in the backbone [3]. Unlearning a source at deployment requires no gradient updates and no access to retained data — operators simply disable the corresponding sinks [4]. The researchers scaled NULLs to Wikipedia’s roughly 6 million articles, treating each article as an independent source [1]. When a single article was unlearned, the model dropped facts unique to that article but preserved information shared with semantically related pages, a result the authors say closely matches retraining from scratch [5]. A separate case study tested the approach on the Harry Potter books; the authors report that NULLs resisted both adversarial extraction and relearning attacks that have reversed other post-hoc unlearning methods [1]. Machine unlearning has drawn widening attention as large language models are deployed in settings governed by privacy regulations, copyright claims, and safety requirements. A recent survey of the field categorizes existing approaches into data-centric, parameter-centric, architecture-centric, and hybrid strategies, and notes that providing formal guarantees of forgetting while scaling to billion-parameter models remains an open challenge [9]. NULLs falls into the architecture-centric category by baking source isolation into the model design rather than applying it after training. On standard downstream benchmarks, the NULLs model matched the performance of a conventional transformer, suggesting that native unlearning does not come at the cost of general language capability [1]. The submission, authored by Gaurav Rohit Ghosal and colleagues, runs to 1,404 KB [1]. The paper argues that source-level unlearning “need not be an afterthought” and can be built directly into LLM training while retaining the benefits of shared representation learning [3].