浙江大学的 SpaceCast-Bench 测试了 VLM 是否能够在场景发生变化后预测空间关系。21 个模型中得分最高的为 58.0%,而人类的得分为 87.2%。Zhejiang University's SpaceCast-Bench tests whether VLMs can predict spatial relations after a scene changes. The best of 21 models scored 58.0% against 87.2% for humans.