AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

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

Guozhi Liu, Weiwei Lin, Tiansheng Huang, Ruichao Mo, Qi Mu, Xiumin Wang, Li Shen

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add sink divergence suppression regularizer during fine-tuningSFT on Llama3-8B-Instruct · BeaverTails, GSM8K, avg over harmful ratiosharmful score: 23.10 -> 8.42Improved
add sink divergence suppression regularizerSFT on Llama3-8B-Instruct · BeaverTails, GSM8K, avg over harmful ratiosfine-tune accuracy: 65.58 -> 68.70Improved
use Surgery instead of best baseline on other harm benchmarksLlama3-8B-Instruct · HarmBenchharmful score: 20.75 -> 9.50Improved
use Surgery on Gemma2-9B-InstructGemma2-9B-Instruct · GSM8Kharmful score: 27.30 -> 8.20Improved
compute sink divergence on RepNoise-Refusal instead of BeaverTails (ablation)Surgery · BeaverTailsharmful score: 8.90 -> 13.20Hurt
increase harmful ratio from 0.05 to 0.2 (ablation)Surgery · GSM8Kharmful score: 8.80 -> 9.30No clear effect
lambda from 0.1 to 300 (ablation)Surgery · GSM8K Llama3harmful score: 25.00 -> 8.90Improved
increase lambda from 300 to 500 (ablation)Surgery · GSM8K Qwen2.5-14Bfine-tune accuracy: 90.80 -> 86.70Hurt
lower learning rate to 5e-7 (ablation)Surgery · GSM8K Llama3harmful score and accuracy: HS 8.90 -> 1.50; FA 68.50 -> 19.40Mixed
raise learning rate to 5e-5 (ablation)Surgery · GSM8K Llama3harmful score: 8.90 -> 23.30Hurt
disable attention heads with positive sink divergence (ablation)fine-tuned model · harmful scoreharmfulnessImproved
disable attention heads with positive sink divergence (ablation)fine-tuned model · Figure 3 (SFT, Llama3-8B)harmful scoreImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy