AI Paper Atlas

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

ICLR 2026 · Large language models › Alignment and preferences · approach: Steering and correcting LLM behavior · 5 citations (OpenAlex)

Read the paper

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 changedWhereMeasuredEffect
add multilingual consistency (MLC) loss to DPOQwen-2.5-7B-Instruct · PKU-SafeRLHF avg safety ratesafety rate: 66.44 -> 95.94Improved
add MLC loss to DPOGemma-2-9B-it · PKU-SafeRLHF avg safety ratesafety rate: 81.52 -> 96.83Improved
add MLC loss to DPOQwen-2.5-7B-Instruct · PKU-SafeRLHF Swahilisafety rate: 7.22 -> 92.78Improved
add MLC loss to DPOQwen-2.5-7B-Instruct · PKU-SafeRLHF cross-language variancevariance: 12.44 -> 0.07Improved
add MLC loss to DPOQwen-2.5-7B-Instruct · MultiJail in-distributionASR: 10.54 -> 0.70Improved
add MLC loss to DPOQwen-2.5-7B-Instruct · MultiJail out-of-distribution (unseen languages)ASR: 3.62 -> 0.51Improved
add MLC loss to DPOQwen-2.5-7B · MMMLU-lite Allaccuracy: 54.59 -> 52.66Hurt
add MLC loss to DPOGemma-2-9B-it · MMMLU-lite Allaccuracy: 49.27 -> 54.61Improved
add MLC loss to DPOQwen-2.5-7B · MMLUaccuracy: 76.22 -> 76.30No clear effect
add MLC loss to DPOQwen-2.5-1.5B · PKU-SafeRLHF avg safety ratesafety rate: 66.50 -> 94.11Improved
add MLC loss to DPOQwen-2.5-3B · PKU-SafeRLHF avg safety ratesafety rate: 63.67 -> 92.45Improved
use autoencoder extractor instead of linear projection (ablation)DPO+MLC · PKU-SafeRLHF avg safety ratesafety rate: 95.94 -> 83.17Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy