The field of visual understanding is moving towards the development of more robust and generalizable multimodal models. Researchers are focusing on improving the capabilities of large language models to understand visual content, including images and videos, and to reason about the physical and social principles that govern the world. This includes the development of new benchmarks and datasets to evaluate the visual knowledge understanding of multimodal models. The use of pre-training pipelines and multi-task training workflows is also becoming increasingly popular to enhance the performance and transferability of these models. Noteworthy papers include VITAL, which proposes a vision-encoder-centered pre-training pipeline for visual quality assessment, and MASS, which introduces a motion-aware spatial-temporal grounding method for physics reasoning and comprehension in vision-language models. VKnowU is also notable for evaluating visual knowledge understanding in multimodal large language models and introducing a new dataset and baseline model to bridge the gap in world-centric visual knowledge.