Research map / Large language models / Instruction and fine-tuning
Superficial Safety Alignment Hypothesis
Jianwei Li, Jung-Eun Kim
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 |
|---|---|---|---|
| freeze all SCU and top 6% of CU during benign finetuning (ablation) | Llama2-7B-Chat finetuning · HEx-PHI gpt4-rate, Alpaca | high harmfulness rate (%): 16.1% -> 7.2% | Improved |
| freeze all CU in addition to all SCU (instead of only top 6% CU) (ablation) | Llama2-7B-Chat finetuning · HEx-PHI gpt4-rate, Alpaca | high harmfulness rate (%): 7.2% -> 4.5% | Improved |
| freeze SCU and CU on Llama3 (first 12 blocks frozen) during Dolly finetuning (ablation) | Llama3-8B-Instruct finetuning · AdvBench keyword, Dolly | attack success rate (%): 61.15% -> 40.58% | Improved |
| pruning the safety critical units (1.3% of units) (ablation) | Llama2-7B-Chat · safety ASR average | ASR (%): 10.0 -> 66.0 | Hurt |
| pruning utility critical units (ablation) | Llama2-7B-Chat · utility average accuracy | accuracy (%): 58.3 -> 42.7 | Hurt |
| pruning redundant units (ablation) | Llama3-8B-Instruct · utility average accuracy | accuracy (%): 61.2 -> 60.8 | No clear effect |
| use parameter-efficient finetuning (LoRA, LLaMA-Adapter, Prefix) instead of full finetuning (ablation) | Llama2-7B-Chat finetuning on Alpaca · HEx-PHI high harmfulness rate | high harmfulness rate (%): 16.1% -> 25.2% (LoRA) | Hurt |
| finetune only the redundant units (20% of parameters) instead of all parameters (ablation) | Llama-7B Alpaca SFT · GSM8K 5-shot | accuracy: 8.8 -> 13.4 | Improved |
| finetune only the redundant units instead of all parameters (ablation) | Llama-7B Alpaca SFT · MMLU | accuracy: 37.8 -> 33.7 | Hurt |
| finetune only the redundant units instead of all parameters (ablation) | Llama-7B Alpaca SFT · MT-bench first turn | score: 2.83 -> 3.5 | Improved |
| freeze SCU and CU during finetuning (utility check) (ablation) | Llama2-7B-Chat Alpaca finetuning · GSM8K | accuracy: 19.79 -> 19.64 | No clear effect |
| use SCU/CU maximal-margin freezing instead of neuron-level safety neuron freezing | Llama2-7B-Chat Alpaca finetuning · HEx-PHI llama3-guard | harmful rate (%): 13.33% -> 12.12% | Improved |
Closest papers by meaning
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.94)
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications (ICML 2024 · similarity 0.93)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.92)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.92)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.92)
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 2025 · similarity 0.92)
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.
