_version_ 1866911391944802304
author Vij, Akriti
Chua, Benjamin
Ramiah, Darshini
Ng, En Qi
Morsidi, Mahran
Gangarapu, Naga Nikshith
Johnson, Sharmini
Wilfred, Vanessa
Kumaran, Vikneswaran
Lee, Wan Sie
Yang, Wenzhuo
Zheng, Yongsen
Black, Bill
Xia, Boming
Sun, Frank
Zhang, Hao
Lu, Qinghua
Ma, Suyu
Liu, Yue
Lo, Chi-kiu
Azadi, Fatemeh
Nejadgholi, Isar
Vajjala, Sowmya
Delaborde, Agnes
Rolin, Nicolas
Seimandi, Tom
Murakami, Akiko
Ishi, Haruto
Sekine, Satoshi
Semitsu, Takayuki
Sasaki, Tasuku
Kinuthia, Angela
Wangari, Jean
Michie, Michael
Kasaon, Stephanie
Baek, Hankyul
Noh, Jaewon
Nam, Kihyuk
Seo, Sang
Shin, Sungpil
Lee, Taewhi
Kim, Yongsu
Newbold-Harrop, Daisy
Wang, Jessica
Ghanem, Mahmoud
Hong, Vy
author_facet Vij, Akriti
Chua, Benjamin
Ramiah, Darshini
Ng, En Qi
Morsidi, Mahran
Gangarapu, Naga Nikshith
Johnson, Sharmini
Wilfred, Vanessa
Kumaran, Vikneswaran
Lee, Wan Sie
Yang, Wenzhuo
Zheng, Yongsen
Black, Bill
Xia, Boming
Sun, Frank
Zhang, Hao
Lu, Qinghua
Ma, Suyu
Liu, Yue
Lo, Chi-kiu
Azadi, Fatemeh
Nejadgholi, Isar
Vajjala, Sowmya
Delaborde, Agnes
Rolin, Nicolas
Seimandi, Tom
Murakami, Akiko
Ishi, Haruto
Sekine, Satoshi
Semitsu, Takayuki
Sasaki, Tasuku
Kinuthia, Angela
Wangari, Jean
Michie, Michael
Kasaon, Stephanie
Baek, Hankyul
Noh, Jaewon
Nam, Kihyuk
Seo, Sang
Shin, Sungpil
Lee, Taewhi
Kim, Yongsu
Newbold-Harrop, Daisy
Wang, Jessica
Ghanem, Mahmoud
Hong, Vy
contents As frontier AI models are deployed globally, it is essential that their behaviour remains safe and reliable across diverse linguistic and cultural contexts. To examine how current model safeguards hold up in such settings, participants from the International Network for Advanced AI Measurement, Evaluation and Science, including representatives from Singapore, Japan, Australia, Canada, the EU, France, Kenya, South Korea and the UK conducted a joint multilingual evaluation exercise. Led by Singapore AISI, two open-weight models were tested across ten languages spanning high and low resourced groups: Cantonese English, Farsi, French, Japanese, Korean, Kiswahili, Malay, Mandarin Chinese and Telugu. Over 6,000 newly translated prompts were evaluated across five harm categories (privacy, non-violent crime, violent crime, intellectual property and jailbreak robustness), using both LLM-as-a-judge and human annotation. The exercise shows how safety behaviours can vary across languages. These include differences in safeguard robustness across languages and harm types and variation in evaluator reliability (LLM-as-judge vs. human review). Further, it also generated methodological insights for improving multilingual safety evaluations, such as the need for culturally contextualised translations, stress-tested evaluator prompts and clearer human annotation guidelines. This work represents an initial step toward a shared framework for multilingual safety testing of advanced AI systems and calls for continued collaboration with the wider research community and industry.
format Preprint
id arxiv_https___arxiv_org_abs_2601_15706
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Improving Methodologies for LLM Evaluations Across Global Languages
Vij, Akriti
Chua, Benjamin
Ramiah, Darshini
Ng, En Qi
Morsidi, Mahran
Gangarapu, Naga Nikshith
Johnson, Sharmini
Wilfred, Vanessa
Kumaran, Vikneswaran
Lee, Wan Sie
Yang, Wenzhuo
Zheng, Yongsen
Black, Bill
Xia, Boming
Sun, Frank
Zhang, Hao
Lu, Qinghua
Ma, Suyu
Liu, Yue
Lo, Chi-kiu
Azadi, Fatemeh
Nejadgholi, Isar
Vajjala, Sowmya
Delaborde, Agnes
Rolin, Nicolas
Seimandi, Tom
Murakami, Akiko
Ishi, Haruto
Sekine, Satoshi
Semitsu, Takayuki
Sasaki, Tasuku
Kinuthia, Angela
Wangari, Jean
Michie, Michael
Kasaon, Stephanie
Baek, Hankyul
Noh, Jaewon
Nam, Kihyuk
Seo, Sang
Shin, Sungpil
Lee, Taewhi
Kim, Yongsu
Newbold-Harrop, Daisy
Wang, Jessica
Ghanem, Mahmoud
Hong, Vy
Artificial Intelligence
As frontier AI models are deployed globally, it is essential that their behaviour remains safe and reliable across diverse linguistic and cultural contexts. To examine how current model safeguards hold up in such settings, participants from the International Network for Advanced AI Measurement, Evaluation and Science, including representatives from Singapore, Japan, Australia, Canada, the EU, France, Kenya, South Korea and the UK conducted a joint multilingual evaluation exercise. Led by Singapore AISI, two open-weight models were tested across ten languages spanning high and low resourced groups: Cantonese English, Farsi, French, Japanese, Korean, Kiswahili, Malay, Mandarin Chinese and Telugu. Over 6,000 newly translated prompts were evaluated across five harm categories (privacy, non-violent crime, violent crime, intellectual property and jailbreak robustness), using both LLM-as-a-judge and human annotation. The exercise shows how safety behaviours can vary across languages. These include differences in safeguard robustness across languages and harm types and variation in evaluator reliability (LLM-as-judge vs. human review). Further, it also generated methodological insights for improving multilingual safety evaluations, such as the need for culturally contextualised translations, stress-tested evaluator prompts and clearer human annotation guidelines. This work represents an initial step toward a shared framework for multilingual safety testing of advanced AI systems and calls for continued collaboration with the wider research community and industry.
title Improving Methodologies for LLM Evaluations Across Global Languages
topic Artificial Intelligence
url https://arxiv.org/abs/2601.15706