Tag
This paper introduces EduArt, an educational-level benchmark with 871 human-authored questions for evaluating art history knowledge and visual reasoning in multimodal LLMs, revealing that single-format benchmarks overestimate model capabilities.