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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add sink divergence suppression regularizer during fine-tuning | SFT on Llama3-8B-Instruct · BeaverTails, GSM8K, avg over harmful ratios | harmful score: 23.10 -> 8.42 | Improved |
| add sink divergence suppression regularizer | SFT on Llama3-8B-Instruct · BeaverTails, GSM8K, avg over harmful ratios | fine-tune accuracy: 65.58 -> 68.70 | Improved |
| use Surgery instead of best baseline on other harm benchmarks | Llama3-8B-Instruct · HarmBench | harmful score: 20.75 -> 9.50 | Improved |
| use Surgery on Gemma2-9B-Instruct | Gemma2-9B-Instruct · GSM8K | harmful score: 27.30 -> 8.20 | Improved |
| compute sink divergence on RepNoise-Refusal instead of BeaverTails (ablation) | Surgery · BeaverTails | harmful score: 8.90 -> 13.20 | Hurt |
| increase harmful ratio from 0.05 to 0.2 (ablation) | Surgery · GSM8K | harmful score: 8.80 -> 9.30 | No clear effect |
| lambda from 0.1 to 300 (ablation) | Surgery · GSM8K Llama3 | harmful score: 25.00 -> 8.90 | Improved |
| increase lambda from 300 to 500 (ablation) | Surgery · GSM8K Qwen2.5-14B | fine-tune accuracy: 90.80 -> 86.70 | Hurt |
| lower learning rate to 5e-7 (ablation) | Surgery · GSM8K Llama3 | harmful score and accuracy: HS 8.90 -> 1.50; FA 68.50 -> 19.40 | Mixed |
| raise learning rate to 5e-5 (ablation) | Surgery · GSM8K Llama3 | harmful score: 8.90 -> 23.30 | Hurt |
| disable attention heads with positive sink divergence (ablation) | fine-tuned model · harmful score | harmfulness | Improved |
| disable attention heads with positive sink divergence (ablation) | fine-tuned model · Figure 3 (SFT, Llama3-8B) | harmful score | Improved |
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.
- AsFT — AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin (AAAI 2026) · better on 3 of 3 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
- Learning and Forgetting Unsafe Examples in Large Language Models (ICML 2024 · similarity 0.89)
- On the Role of Attention Heads in Large Language Model Safety (ICLR 2025 · similarity 0.89)
- Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning Perturbation (NeurIPS 2025 · similarity 0.88)
- Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning Attack (ICML 2025 · similarity 0.87)
- Do as I do (Safely): Mitigating Task-Specific Fine-tuning Risks in Large Language Models (ICLR 2025 · similarity 0.87)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.87)
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.
