Who Brought Easter Eggs to Eid? Auditing Cultural Translation of Math Word Problems Across Diverse Languages and Regions

This study analyzes the cultural adaptation of math word problems by three large language models: Claude Opus 4, GPT-4.1, and Gemini 2.5 Pro, across multiple languages including Bengali, Hindi, Punjabi, Urdu, Sindhi, Italian, and Sicilian. The authors found that while models agree on transformation types in 62.5% of cases, they only align on specific substitutions 33.5% of the time, leading to significant discrepancies in cultural representation and potential misattributions of regional context. This highlights the importance for practitioners to critically evaluate model outputs to ensure cultural accuracy and diversity in educational applications, as adaptations often prioritize superficial markers over deeper contextual integrity.