Research map / Large language models / Language, safety and interpretability
Who Transfers Safety? Identifying and Targeting Cross-Lingual Shared Safety Neurons
Xianhui Zhang, Chengyu Xie, Linxia Zhu, Yonghui Yang, Weixiang Zhao, Zifeng Cheng, Cong Wang, Fei Shen, Tat-Seng Chua
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 English-MS-neuron-only SS-Neuron expansion tuning | Gemma2-9B-it · MultiJail | avg ASR (%): 10.93 -> 1.59 | Improved |
| add SS-Neuron expansion tuning | Llama3.1-8B-it · AdvBench-x | ASR (%): 30.22 -> 0.22 | Improved |
| add NHR translated parallel data to HR-only training (ablation) | SS-Neuron expansion · MultiJail (Gemma2-9B-it) | ASR (%): 23.49 -> 1.59 | Improved |
| add NHR translated parallel data to HR-only training (ablation) | SS-Neuron expansion · AdvBench-x (Qwen3-8B) | ASR (%): 12.20 -> 1.13 | Improved |
| select English MS-Neurons instead of random neurons (ablation) | SS-Neuron expansion · MultiJail (Gemma2-9B-it) | ASR (%): 5.12 -> 1.59 | Improved |
| select English MS-Neurons instead of random neurons (ablation) | SS-Neuron expansion · AdvBench-x (Llama3.1-8B-it) | ASR (%): 5.19 -> 0.22 | Improved |
| mask MS-Neurons (causal check) | Qwen3-8B · AdvBench-x | ASR (%): 15.60 -> 41.46 | Hurt |
| mask randomly chosen neurons (control) | Qwen3-8B · AdvBench-x | ASR (%): 15.60 -> 15.38 | No clear effect |
| mask SS-Neurons on NHR languages | Qwen3-8B · MultiJail | ASR (%): 22.12 -> 33.97 | Hurt |
| use sparse neuron tuning instead of LoRA | Gemma2-9B-it · MultiJail | ASR (%): 33.06 -> 1.59 | Improved |
| use neuron tuning instead of full training | Gemma2-9B-it · MMLU | accuracy: 12.88 -> 79.36 | Improved |
| use neuron tuning instead of LoRA | Qwen3-8B · MMLU | accuracy: 78.40 -> 84.08 | 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.
- SimPO — SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024) · better on 2 of 2 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
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.90)
- A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models (NeurIPS 2026 · similarity 0.90)
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications (ICML 2024 · similarity 0.87)
- Refusal Direction is Universal Across Safety-Aligned Languages (NeurIPS 2025 · similarity 0.87)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.86)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.86)
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.
