The field of multimodal modeling is rapidly advancing, with a focus on developing more efficient and effective models for tasks such as image generation, understanding, and manipulation. Recent research has explored the use of novel architectures, such as two-end-separated and middle-shared designs, to improve the performance of unified multimodal models. Additionally, there is a growing interest in incorporating physical and biological priors into generative models, particularly in the context of protein structure prediction and generation. Notable papers in this area include SpecMER, which introduces a speculative decoding framework for fast protein generation, and UniAlignment, which proposes a unified multimodal generation framework for tasks such as image understanding and manipulation. Other noteworthy papers include HieraTok, which presents a multi-scale visual tokenizer for image reconstruction and generation, and MarS-FM, which introduces a generative model for molecular dynamics simulation. Overall, the field is moving towards more integrated and physically-informed models that can effectively capture the complexities of multimodal data.