Vision-Language Models

Exposing language shortcuts in 3D reasoning benchmarks and introducing Real-3DQA to test whether models actually use visual geometry.