# Surgery: Mitigating Harmful Fine-Tuning for Large Language Models via Attention Sink

arXiv: https://arxiv.org/abs/2602.05228 · HTML: https://mlsanity.com/paper/2602.05228/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 132 of its 167 references that are in the map.
- First posted on arXiv 5 Feb 2026.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Safety alignment and jailbreak attacks](https://mlsanity.com/alignment/safety-alignment-and-jailbreak-attacks/)
- LLM safety and security / [Safety alignment, refusal and guardrails](https://mlsanity.com/safety/safety-alignment-refusal-and-guardrails/)

## Builds on (most cited first, 30 of 132)

- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [Character-level Convolutional Networks for Text Classification](https://mlsanity.com/paper/1509.01626/) (2015)
- [Efficient Streaming Language Models with Attention Sinks](https://mlsanity.com/paper/2309.17453/) (2023)
- [HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal](https://mlsanity.com/paper/2402.04249/) (2024)
- [Inference-Time Intervention: Eliciting Truthful Answers from a Language Model](https://mlsanity.com/paper/2306.03341/) (2023)
- [Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!](https://mlsanity.com/paper/2310.03693/) (2023)
- [Refusal in Language Models Is Mediated by a Single Direction](https://mlsanity.com/paper/2406.11717/) (2024)
- [BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset](https://mlsanity.com/paper/2307.04657/) (2023)
- [LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models](https://mlsanity.com/paper/2311.17043/) (2023)
- [Safety Alignment Should Be Made More Than Just a Few Tokens Deep](https://mlsanity.com/paper/2406.05946/) (2024)
- [Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions](https://mlsanity.com/paper/2309.07875/) (2023)
- [Shadow Alignment: The Ease of Subverting Safely-Aligned Language Models](https://mlsanity.com/paper/2310.02949/) (2023)
- [Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free](https://mlsanity.com/paper/2505.06708/) (2025)
- [SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal](https://mlsanity.com/paper/2406.14598/) (2024)
- [When Attention Sink Emerges in Language Models: An Empirical View](https://mlsanity.com/paper/2410.10781/) (2024)
- [LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B](https://mlsanity.com/paper/2310.20624/) (2023)
- [Removing RLHF Protections in GPT-4 via Fine-Tuning](https://mlsanity.com/paper/2311.05553/) (2023)
- [Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models](https://mlsanity.com/paper/2402.02207/) (2024)
- [Tamper-Resistant Safeguards for Open-Weight LLMs](https://mlsanity.com/paper/2408.00761/) (2024)
- [Safety Layers in Aligned Large Language Models: The Key to LLM Security](https://mlsanity.com/paper/2408.17003/) (2024)
- [See What You Are Told: Visual Attention Sink in Large Multimodal Models](https://mlsanity.com/paper/2503.03321/) (2025)
- [Why do LLMs attend to the first token?](https://mlsanity.com/paper/2504.02732/) (2025)
- [Safe LoRA: the Silver Lining of Reducing Safety Risks when Fine-tuning Large Language Models](https://mlsanity.com/paper/2405.16833/) (2024)
- [What is in Your Safe Data? Identifying Benign Data that Breaks Safety](https://mlsanity.com/paper/2404.01099/) (2024)
- [Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack](https://mlsanity.com/paper/2402.01109/) (2024)
- [Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic](https://mlsanity.com/paper/2402.11746/) (2024)
- [Fine-tuning can cripple your foundation model; preserving features may be the solution](https://mlsanity.com/paper/2308.13320/) (2023)
- [Keeping LLMs Aligned After Fine-tuning: The Crucial Role of Prompt Templates](https://mlsanity.com/paper/2402.18540/) (2024)
- [Representation Noising: A Defence Mechanism Against Harmful Finetuning](https://mlsanity.com/paper/2405.14577/) (2024)
- [Lisa: Lazy Safety Alignment for Large Language Models against Harmful Fine-tuning Attack](https://mlsanity.com/paper/2405.18641/) (2024)

## Built on by (most cited first, 4 of 4)

- [Attention Sinks Are Provably Necessary in Softmax Transformers: Evidence from Trigger-Conditional Tasks](https://mlsanity.com/paper/2603.11487/) (2026)
- [Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation](https://mlsanity.com/paper/2604.10098/) (2026)
- [Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators in Transformers](https://mlsanity.com/paper/2603.17771/) (2026)
- [SafeTune: A Unified Faithful Library for Auditing and Repairing Safety Drift in Fine-Tuned LLMs](https://mlsanity.com/paper/2609.22153/) (2026)

## Abstract (arXiv)

Harmful fine-tuning can invalidate safety alignment of large language models, exposing significant safety risks. In this paper, we utilize the attention sink mechanism to mitigate harmful fine-tuning. Specifically, we first measure a statistic named sink divergence for each attention head and observe that different attention heads exhibit two different signs of sink divergence. To understand its safety implications, we conduct experiments and find that the number of attention heads of positive sink divergence increases along with the increase of the model's harmfulness when undergoing harmful fine-tuning. Based on this finding, we propose a separable sink divergence hypothesis -- attention heads associating with learning harmful patterns during fine-tuning are separable by their sign of sink divergence. Based on the hypothesis, we propose a fine-tuning-stage defense, dubbed Surgery. Surgery utilizes a regularizer for sink divergence suppression, which steers attention heads toward the negative sink divergence group, thereby reducing the model's tendency to learn and amplify harmful patterns. Extensive experiments demonstrate that Surgery improves defense performance by 5.90%, 11.25%, and 9.55% on the BeaverTails, HarmBench, and SorryBench benchmarks, respectively. Source code is available on https://github.com/Lslland/Surgery.