Teaching vision models to reason across multiple viewpoints

Multimodal language models struggle to reason consistently about objects and scenes viewed from multiple angles. CrossView Suite addresses three concrete gaps: lack of training data, absence of evaluation benchmarks, and no explicit mechanisms for matching objects across views. The suite includes CrossViewSet (1.6M annotated samples), CrossViewBench (scene-disjoint evaluation benchmark), and CrossViewer (a three-stage framework using spatial region tokenization, cross-view object alignment, and multi-view feature fusion). Experiments confirm that large-scale training data, systematic evaluation, and explicit alignment all significantly improve cross-view spatial reasoning. Code and models are released.