Despite the rapid progress of Multimodal Large Language Models (MLLMs) in 2D vision-language tasks, robust multi-view spatial reasoning remains a fundamental…
机构:中科大
来源:arXiv 2609.07258 | AI4Papers 论文推荐平台