Lune

Abstract

We reveal the theoretical foundations of techniques for editing large language models, and present new methods which can do so without requiring retraining. Our theoretical insights show that a single metric (a measure of the intrinsic dimension of the model's features) can be used to assess a model's editability and reveals its previously unrecognised susceptibility to malicious stealth attacks. This metric is fundamental to predicting the success of a variety of editing approaches, and reveals new bridges between disparate families of editing methods. We collectively refer to these as stealth editing methods, because they directly update a model's weights to specify its response to specific known hallucinating prompts without affecting other model behaviour. By carefully applying our theoretical insights, we are able to introduce a new jet-pack network block which is optimised for highly selective model editing, uses only standard network operations, and can be inserted into existing networks. We also reveal the vulnerability of language models to stealth attacks: a small change to a model's weights which fixes its response to a single attacker-chosen prompt. Stealth attacks are computationally simple, do not require access to or knowledge of the model's training data, and therefore represent a potent yet previously unrecognised threat to redistributed foundation models. Extensive experimental results illustrate and support our methods and their theoretical underpinnings. Demos and source code are available at https://github.com/qinghua-zhou/stealth-edits.

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

3 top-tier papers have cited it since. Know what they found before your reviewers do.

Questions to start from

Lune can draw on

the full paper · the top-tier papers that cite it · related work

Get the answer

Free to start. No credit card required.

Lineage

  • Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors cites Locating and Editing Factual Associations in GPT
  • Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors cites Fast Model Editing at Scale
  • Locating and Editing Factual Associations in GPT cites Fast Model Editing at Scale
  • Resolving Lexical Bias in Model Editing cites Locating and Editing Factual Associations in GPT
  • Resolving Lexical Bias in Model Editing cites Fast Model Editing at Scale
Builds onCited byLeft to right by year; the more cited, the higher.

Cited by top-tier papers3

  1. Staining and Locking Computer Vision Models Without RetrainingOliver J. Sutton, Qinghua Zhou, George Leete, Alexander N. Gorban et al.ICCV 2025 · 2 citations
  2. Resolving Lexical Bias in Model EditingHammad Rizwan, Domenic Rosati, Ga Wu, Hassan SajjadICML 2025
  3. ShadeEdit: A Utility-Preserving and Defense-Evasive Knowledge Manipulation Attack in Federated LLMsXu Zhang, Hangcheng Liu, Shangwei Guo, Shudong Zhang et al.AAAI 2026

Builds on12

  1. Training language models to follow instructions with human feedbackLong Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida et al.NeurIPS 2022 · 24,707 citations
  2. Retrieval-Augmented Generation for Knowledge-Intensive NLP TasksPatrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni et al.NeurIPS 2020 · 19,162 citations
  3. Locating and Editing Factual Associations in GPTKevin Meng, David Bau, Alex Andonian, Yonatan BelinkovNeurIPS 2022 · 3,415 citations
  4. Fast Model Editing at ScaleEric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn et al.ICLR 2022 · 527 citations
  5. Aging with GRACE: Lifelong Model Editing with Discrete Key-Value AdaptorsTom Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim et al.NeurIPS 2023 · 349 citations

Related papers

  1. BadEdit: Backdooring Large Language Models by Model EditingYanzhou Li, Tianlin Li, Kangjie Chen, Jian Zhang et al.ICLR 2024 · 116 citations
  2. Concept-ROT: Poisoning Concepts in Large Language Models with Model EditingKeltin Grimes, Marco Christiani, David Shriver, Marissa Catherine ConnorICLR 2025
  3. Injecting Universal Jailbreak Backdoors into LLMs in MinutesZhuowei Chen, Qiannan Zhang, Shichao PeiICLR 2025
  4. Tracing and Reversing Edits in LLMsPaul Youssef, Zhixue Zhao, Christin Seifert, Jörg SchlöttererICLR 2026 · 7 citations
  5. Compiling Activation Steering into Weights via Null-Space Constraints for Stealthy BackdoorsRui Yin, Tianxu Han, Naen Xu, Changjiang Li et al.ACL 2026
Dusk over the sea between two cliffs drawn in fine vertical lines