1 paper
Ruoxuan Zhang, Qiyun Zheng, Siyu Wu +13
Vision-Language Models (VLMs) are widely used for visual understanding, yet current evaluation protocols fail to assess whether these capabilities are grounded in physical reasonin…