Meta has introduced UniBench, a new benchmark aimed at enhancing visual reasoning capabilities in vision-language models (VLMs). The initiative emphasizes that merely scaling existing models provides limited advantages for reasoning tasks. Notably, leading VLMs still encounter difficulties with basic digit recognition and counting tasks, such as those presented by the MNIST dataset. Additionally, Alibaba has unveiled mPLUG-Owl3, a model designed for understanding long image sequences and videos. This model employs innovative hyper attention blocks to facilitate efficient semantic alignment between vision and language. Both developments reflect a growing focus on improving the reasoning and understanding capabilities of multimodal large language models.