Browse Teach Curate

Benchmarking and Improving Cultural Understanding in Vision-Language Models

How to measure and reduce the Western-centric cultural bias of vision-language models: building a large-scale multimodal benchmark of culturally specific symbols, gestures, and artifacts spanning many countries, using it to expose systematic regional performance gaps across many models, and fine-tuning on culturally rich image-text data to raise cultural understanding while preserving general capabilities and generalizing across cultures, continents, and datasets.

Demerzel Text

CultureVLM: Characterizing and Improving Cultural Understanding of Vision-Language Models for over

Constructs CultureVerse, a large-scale multimodal benchmark of 19,682 cultural concepts and 228,053 image-text samples spanning 188 countries/regions and 15 cultural categories with three question ty…