Cross-Cultural Personalization Benchmarking for Multimodal Large Language Models

Building on Nayak et al.’s CulturalVQA (Source 1) which reveals cultural disparities in vision-language models’ understanding, this research proposes a personalized benchmarking framework that explicitly models cultural user profiles to evaluate how well multimodal LLMs personalize outputs across diverse cultural dimensions such as traditions, values, and communication styles. Unlike prior benchmarks that assess static cultural knowledge, this framework dynamically adapts user contexts to simulate real-world multicultural interactions, measuring model robustness and adaptability for personalized cultural understanding. This research uniquely synthesizes personalized user modeling (Sources 3, 4 from Heuristic: Create novel syntheses) with multimodal cultural evaluation, addressing an underexplored intersection. The framework can guide development of culturally aware AI assistants and personalized education tools, reducing bias and improving global usability.