The Multilingual Alignment Prism Aligning Global And Local Preferences To Reduce Harm 2024 06 08

Language-model safety tests need to account for whose language and preferences they represent. Cohere's Multilingual Alignment Prism abstract describes human-annotated prompts designed to expose harmful responses, distinguishing broadly shared harms from locally specific ones. The researchers use these prompts to examine alignment approaches across six languages and report strong alignment results with little degradation in general performance. The contribution is a way to evaluate safety across differing cultural and linguistic preferences, rather than treating English-focused harm reduction as sufficient evidence for every audience.