Research map / Large language models / Alignment and preferences
Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment
Yuyan Bu, Xiaohao Liu, ZhaoXing Ren, Yaodong Yang, Juntao Dai
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 multilingual consistency (MLC) loss to DPO | Qwen-2.5-7B-Instruct · PKU-SafeRLHF avg safety rate | safety rate: 66.44 -> 95.94 | Improved |
| add MLC loss to DPO | Gemma-2-9B-it · PKU-SafeRLHF avg safety rate | safety rate: 81.52 -> 96.83 | Improved |
| add MLC loss to DPO | Qwen-2.5-7B-Instruct · PKU-SafeRLHF Swahili | safety rate: 7.22 -> 92.78 | Improved |
| add MLC loss to DPO | Qwen-2.5-7B-Instruct · PKU-SafeRLHF cross-language variance | variance: 12.44 -> 0.07 | Improved |
| add MLC loss to DPO | Qwen-2.5-7B-Instruct · MultiJail in-distribution | ASR: 10.54 -> 0.70 | Improved |
| add MLC loss to DPO | Qwen-2.5-7B-Instruct · MultiJail out-of-distribution (unseen languages) | ASR: 3.62 -> 0.51 | Improved |
| add MLC loss to DPO | Qwen-2.5-7B · MMMLU-lite All | accuracy: 54.59 -> 52.66 | Hurt |
| add MLC loss to DPO | Gemma-2-9B-it · MMMLU-lite All | accuracy: 49.27 -> 54.61 | Improved |
| add MLC loss to DPO | Qwen-2.5-7B · MMLU | accuracy: 76.22 -> 76.30 | No clear effect |
| add MLC loss to DPO | Qwen-2.5-1.5B · PKU-SafeRLHF avg safety rate | safety rate: 66.50 -> 94.11 | Improved |
| add MLC loss to DPO | Qwen-2.5-3B · PKU-SafeRLHF avg safety rate | safety rate: 63.67 -> 92.45 | Improved |
| use autoencoder extractor instead of linear projection (ablation) | DPO+MLC · PKU-SafeRLHF avg safety rate | safety rate: 95.94 -> 83.17 | Hurt |
Closest papers by meaning
- LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations (NeurIPS 2026 · similarity 0.89)
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.89)
- Alignment at Pre-training! Towards Native Alignment for Arabic LLMs (NeurIPS 2024 · similarity 0.89)
- Meta-Learning Preferences for Multilingual LLM Alignment (NeurIPS 2026 · similarity 0.89)
- Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples (ICML 2025 · similarity 0.88)
- L3Ms — Lagrange Large Language Models (ICLR 2025 · similarity 0.88)
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.
