1 paper · 1 filter
Tengyue Wang, Kang An, Chenxu Du +8
Despite significant progress in general visual question answering and cross-modal understanding, multimodal large language models still face a pronounced gap in evaluation for comp…