How I Aligned Your Model: Scaling Up Multilingual Safety Training with Templates
AuthorsMax Neuwinger, Hanna Yukhymenko, Imanol Schlag
AffiliationsETH Zurich
This paper uses culturally informed templates to expand safety training and testing for language models across 60 languages.
Key results
Validated prompts across 44 languages.
Evaluation prompts spanning 60 languages.
HarmBench languages; the WildGuardMix training track covers 44 languages.
HarmBench attack success, compared with 0.0% on the original English prompts.
Approximate share of failures found in only one arm that occur only with culturally localized entities.
Multilingual HarmBench attack success, down from 21.5% before fine-tuning.
What the paper found
This paper introduces MSafeTemp, a template-based pipeline for expanding a small set of harmful prompts into multilingual safety-training and evaluation data. It extracts typed slots for harm-bearing terms, translates the surrounding prompt while preserving those slots, then generates generic harmful variants, culturally localized variants from language-specific cultural cards, and matched benign controls; separate models validate the outputs. The resulting corpus contains 1.84M validated WildGuardMix prompts for training across 44 languages and 216,486 HarmBench evaluation prompts across 60 languages. Evaluations show why English-only safety scores can mislead: on HarmBench, Olmo-3.1-32B-Instruct’s attack success rate rises from 0.0% on English seeds to 13.7% on translated prompts. Cultural localization exposes a different set of weaknesses: about 45% of failures found in only one of the translated or localized arms occur only with local entities. Fine-tuning experiments on Olmo-3-7B reduce multilingual HarmBench attack success from 21.5% to 2.9% after supervised fine-tuning followed by direct preference optimization, though helpfulness and general capabilities remain part of the tradeoff. The pipeline uses Gemma-4 for generation and OpenAI’s GPT-OSS models for judging; the evaluated model panel also includes Meta’s Llama and Google’s Gemma. Overall, the work shows that multilingual safety needs culturally grounded testing and training, not translation alone.
Original abstract
Safety alignment for Large Language Models is heavily focused on English with limited multilingual representation, particularly damaging low-resource language performance. Related open training datasets are often limited in size or do not follow region-localized harmful topics to broaden the scope of alignment beyond cultures. We propose a scalable template-based augmentation pipeline to produce massive multilingual refusal and benign corpora, extracting safe templates from harmful prompts, translating them and then filling in the slots with culturally authentic entities, both harmful and benign. As a result, we release MSafeTemp, a validated corpus spanning 60 languages useful for multilingual safety alignment and evaluation. Using the dataset, we show that English-only evaluation misjudges multilingual safety: for several models, attacks that fail in English succeed once translated, while culturally localized variants expose further failures that translation alone misses. We also confirm that augmenting safety training data through templates allows for higher diversity and quality of alignment, resulting in improved safety performance beyond English. MSafeTemp and the respective pipeline are aimed at facilitating research and development of safer and more responsible models for less-represented cultures.
Read the original paperMore in AI Safety
Browse all 54 papers →Is Forgetting Unified? Exploring Concept Unlearning in Unified Multimodal Models
Zhongkai Xue, Junkai Chen, Mingcong Lei, Rumo Zhang, Yixiao Liu, Zhimin Sun, Shu Wu, Qiang Liu, Lihua Zhang, Liang Wang
UniForget aims to make unified multimodal models both stop generating and stop recognizing selected concepts while preserving other capabilities.
LEASH: Stay Safe in Agentic Robotics
Qinfeng Zhu, Weiguang Zhao, Ke Sun, Xihao Guo, Yunxi Jiang, Rui Zhang, Lei Fan, Kaizhu Huang
RoboRisk tests how safely AI agents control robots, while LEASH adds checks intended to prevent dangerous plans and actions.
"Shall We Approve?" Consequence-Aware Approval for Autonomous Coding Agents
Weiheng Wu, Junkai Chen, Zhensu Sun, Jieke Shi, Yunbo Lyu, David Lo
Caper tests what a coding agent’s command actually does in a sandbox before approving it, aiming to block harmful actions without stopping safe ones.