Research map / Large language models / Instruction and fine-tuning
AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin
Shuo Yang, Qihui Zhang, Yuyang Liu, Yue Huang, Xiaojun Jia, Kun-Peng Ning, Jia-Yu Yao, Jigang Wang, Dai Hailiang, Yibing Song, Li Yuan
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 orthogonal-to-alignment-direction regularizer (AsFT) to SFT | Llama-2-7B-Chat LoRA fine-tuning on AGNEWS · AGNEWS (avg over poison ratios) | harmful score: 21.52 -> 4.08 | Improved |
| add AsFT regularizer | SFT · AGNEWS (avg over poison ratios) | fine-tuning accuracy: 83.26 -> 83.78 | No clear effect |
| add AsFT regularizer | SFT · Llama-3-8B (AGNEWS) | harmful score: 73.60 -> 15.20 | Improved |
| add AsFT regularizer | SFT · Different fine-tuning datasets (avg) | harmful score: 35.50 -> 6.90 | Improved |
| add AsFT regularizer | SFT · GSM8K fine-tuning (Llama-2-7B) | fine-tuning accuracy: 23.80 -> 26.00 | Improved |
| AsFT instead of Safe LoRA post-hoc projection | Llama-2-7B-Chat · AGNEWS (avg over poison ratios) | fine-tuning accuracy: 80.98 -> 83.78 | Improved |
| penalize updates orthogonal to alignment direction (d_harm-perp) with larger lambda (ablation) | AsFT · AGNEWS | harmful score | Improved |
| increase lambda beyond 10 (ablation) | AsFT · AGNEWS | fine-tuning accuracy | Hurt |
| restrict updates along d_aligned instead of its orthogonal complement (ablation) | AsFT · AGNEWS | harmful score | No clear effect |
| derive harmful directions from harmful data instead of the aligned-minus-base weight difference (AsFTAlt) (ablation) | AsFT · AGNEWS (avg over sample numbers) | harmful score: 14.80 -> 9.04 (vs SFT); AsFT 3.20 | Mixed |
| use AsFT over SafeInstr/BEA across learning rates (ablation) | fine-tuning defense · learning rate sweep | harmful score | Improved |
Later papers that compared with this method
3 later accepted papers report AsFT as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- A Guardrail for Safety Preservation: When Safety-Sensitive Subspace Meets Harmful-Resistant Null-Space (ICLR 2026 · better on 9 of 9 shared results)
- Toward Safe Quantization-Aware Fine-tuning: Understanding and Mitigating Safety Alignment Degradation (ICML 2026 · better on 3 of 4 shared results)
- Surgery: Mitigating Harmful Fine-Tuning for Large Language Models via Attention Sink (ICML 2026 · better on 3 of 3 shared results)
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.
- Safe LoRA — Safe LoRA: The Silver Lining of Reducing Safety Risks when Finetuning Large Language Models (NeurIPS 2024) · better on 5 of 5 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
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.94)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.94)
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.92)
- Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning Attack (ICML 2025 · similarity 0.92)
- Do as I do (Safely): Mitigating Task-Specific Fine-tuning Risks in Large Language Models (ICLR 2025 · similarity 0.92)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.91)
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.
