Research map / Large language models / Language, safety and interpretability
Multilingual Safety Alignment via Representation-Space Separability
Dan Shi, Zhuowen Han, Deyi Xiong
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 |
|---|---|---|---|
| use SMO instead of SimPO for multilingual safety alignment | LLaMA-3.1-8B-Instruct · MultiJail avg ASR | attack success rate (%): 5.87 -> 1.53 | Improved |
| add SMG example-wise modulation of the SimPO loss (ablation) | SMO on LLaMA-3.1-8B · MultiJail avg ASR | attack success rate (%): 3.49 -> 1.53 | Improved |
| add SMG example-wise modulation of the SimPO loss (ablation) | SMO on LLaMA-3.1-8B · AdvBench-X avg ASR | attack success rate (%): 2.53 -> 0.29 | Improved |
| add SMG example-wise modulation of the SimPO loss (ablation) | SMO on Qwen-2.5-7B · MultiJail avg ASR | attack success rate (%): 29.42 -> 21.22 | Improved |
| add SMG example-wise modulation of the SimPO loss (ablation) | SMO on Qwen-2.5-7B · AdvBench-X avg ASR | attack success rate (%): 16.41 -> 9.97 | Improved |
| add margin loss term (ablation) | SMO on LLaMA-3.1-8B · MultiJail and AdvBench-X ASR | attack success rate (%): 1.32 -> 1.53 (MultiJail) | No clear effect |
| add margin loss term (ablation) | SMO on LLaMA-3.1-8B · MGSM accuracy | accuracy (%): 24.47 -> 25.24 | Improved |
| use Chinese instead of English as anchor language (ablation) | SMO on Qwen2.5-7B-Instruct · MultiJail avg ASR | attack success rate (%): 21.22 -> 20.21 | No clear effect |
| train with SMO | LLaMA-3.1-8B-Instruct · M-MMLU avg accuracy | accuracy (%): 51.63 -> 51.39 | No clear effect |
| train with SMO | LLaMA-3.1-8B-Instruct · MGSM multilingual avg | accuracy (%): 26.49 -> 23.07 | Hurt |
| train with SMO | Qwen-2.5-7B-Instruct · MGSM En | accuracy (%): 70.40 -> 73.20 | Improved |
| train with SMO | LLaMA-3.1-8B-Instruct · harmful-harmless L2 distance on training data (avg) | L2 distance: 23.39 -> 25.69 | 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 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
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.92)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.90)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.90)
- Safety Depth in Large Language Models: A Markov Chain Perspective (NeurIPS 2025 · similarity 0.90)
- Multilingual Jailbreak Challenges in Large Language Models (ICLR 2024 · similarity 0.90)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.90)
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.
