Culturally-Adapted Red-Teaming Across East and Southeast Asian Contexts: A Methodological and Comparative Analysis

The article presents a methodological analysis of culturally-adapted (CA) red-teaming for multilingual safety evaluation of large language models (LLMs), specifically in East and Southeast Asian contexts. It compares direct translation (DT) and CA datasets for Korean, Japanese, Thai, and Khmer, revealing that CA prompts improve attack success rates (mean +9.3 percentage points) and provide significantly more culturally relevant evaluations, with CA scores averaging 2.51 compared to DT scores of 0.17. This highlights the importance of incorporating cultural context into safety benchmarks for LLMs to accurately assess risks and ensure robust performance in diverse environments.