Scientists have revealed a hierarchical structure for information flow within models, which can significantly enhance the transparency of multi-modal AI systems. Zhang Zhi, a researcher, discovered that when multi-modal large models perform tasks, they process information in a structured manner, from lower layers to higher layers. Specifically, the first step involves transmitting comprehensive visual information from the entire image to the corresponding linguistic representation. In the second step, the model sends specific visual information related to the question to the linguistic representation. Lastly, the model channels the synthesized multi-modal information into the final position in the input sequence to aid in generating the final prediction. One notable finding is that the initial answers generated by the model are often in small written form, and the first character is then systematically converted into larger writing. This indicates that the model's content interpretation (semantic processing) and formatting (syntactic processing) are handled in distinct stages. In this study, Zhang used verified interpretable tools (attention force fusion), ensuring the method’s reliability and avoiding the redundancy of re-verification. The research not only deepens our understanding of how multi-modal large models internally process information but also offers theoretical guidance for future improvements in model architecture and cross-modal information integration. Firstly, in terms of model operational efficiency, the study identified key layers for visual-linguistic information fusion, which can be leveraged to optimize model architectures, reduce redundant computation, and speed up inference processes, particularly in tasks like Visual Question Answering (VQA) and image caption generation. Secondly, in model editing, the findings provide insights into how different modal information operates at different layers. This aids in developing more precise methods for multi-modal information pruning, enabling models to better adapt to specific tasks or scenarios, such as medical image analysis, autonomous driving, and intelligent monitoring. Lastly, in interpretability, the research uncovered the hierarchical structure of internal information flow within the model, enhancing the transparency of multi-modal AI systems. This theoretical support helps develop more controllable and credible AI models, especially in fields requiring rigorous scrutiny, such as law, finance, and medical AI. Overall, this study opens new avenues for refining and optimizing multi-modal AI systems, making them more efficient, adaptable, and transparent. This advancement could have significant implications for both academic and industrial applications.