![[2602.16660] Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16660] Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment
Summary
The paper presents a method for enhancing multilingual safety alignment in large language models (LLMs) using a resource-efficient Multi-Lingual Consistency (MLC) loss, which allows for simultaneous alignment across multiple languages without extensive supervision.
Why It Matters
As LLMs are increasingly deployed in diverse linguistic contexts, ensuring their safety and reliability across languages is crucial. This research addresses the challenge of multilingual alignment, proposing a scalable solution that minimizes resource requirements while improving model performance in low-resource languages.
Key Takeaways
- Introduces a Multi-Lingual Consistency (MLC) loss for LLMs.
- Enables multilingual safety alignment without extensive supervision.
- Demonstrates effectiveness across various model architectures.
- Improves cross-lingual generalization with minimal impact on utility.
- Offers a practical solution for multilingual consistency in AI applications.
Computer Science > Computation and Language arXiv:2602.16660 (cs) [Submitted on 18 Feb 2026] Title:Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment Authors:Yuyan Bu, Xiaohao Liu, ZhaoXing Ren, Yaodong Yang, Juntao Dai View a PDF of the paper titled Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment, by Yuyan Bu and 4 other authors View PDF HTML (experimental) Abstract:The widespread deployment of large language models (LLMs) across linguistic communities necessitates reliable multilingual safety alignment. However, recent efforts to extend alignment to other languages often require substantial resources, either through large-scale, high-quality supervision in the target language or through pairwise alignment with high-resource languages, which limits scalability. In this work, we propose a resource-efficient method for improving multilingual safety alignment. We introduce a plug-and-play Multi-Lingual Consistency (MLC) loss that can be integrated into existing monolingual alignment pipelines. By improving collinearity between multilingual representation vectors, our method encourages directional consistency at the multilingual semantic level in a single update. This allows simultaneous alignment across multiple languages using only multilingual prompt variants without requiring additional response-level supervision in low-resource languages. We validate the proposed method across di...
