MUSE: Benchmarking Large Vision-Language Models on Multi-Modal Understanding in Situated Education
MUSE tests whether vision-language models can read art for classroom use, not just identify what is in an image.
The benchmark covers twelve tasks across perception, meaning, affect, culture, and composition. Its image set is curated around Singaporean and Southeast Asian multicultural contexts as well as Western art traditions. The authors report large performance gaps between models, with affective interpretation and compositional reasoning standing out as weak points. ArXiv · AI/CL/LG's note
The benchmark covers twelve tasks across perception, meaning, affect, culture, and composition. Its image set is curated around Singaporean and Southeast Asian multicultural contexts as well as Western art traditions. The authors report large performance gaps between models, with affective interpretation and compositional reasoning standing out as weak points. ArXiv · AI/CL/LG's note
score 4