Researchers have created a benchmark called VGI-BENCH to evaluate the visual intelligence of video generation models. The benchmark consists of 27 tasks and 810 instances, organized by task domains and skill tags. It tests whether these models can perform visually grounded reasoning tasks, such as understanding scenes and objects in videos. Current generative systems are found to be limited in their ability to solve these tasks, with the strongest model achieving only 51% accuracy under the benchmark's evaluation criteria. --- Why it matters: This matters because it highlights the limitations of current video generation models in performing visually grounded reasoning tasks, which is essential for applications such as video editing and content creation. Understanding these limitations can help researchers develop more advanced models that can better understand visual information. Source: https://arxiv.org/abs/2608.19583