A Control-Theoretic MCP Framework for MLLMs’ Efficiency and Interpretability

Aiming at the computational inefficiency and insufficient interpretability of multimodal large language models (MLLMs) in complex tasks such as multi-round reasoning and medical diagnosis, this paper proposes the MCP (Model-Controller-Presenter) three-layer collaborative framework. By decoupling MLLMs into reasoning, generation, and retrieval sub-modules, integrating a reinforcement learning (RL)-driven dynamic routing algorithm, and designing a task adaptation mechanism, the framework realizes the systematic integration of control theory and MLLM dynamic reasoning for the first time. Experiments on cross-modal benchmark datasets (GLUE, COCO, ScienceQA) show that compared with baseline models (LLaMA-2 7B, GPT-3.5, etc.), MCP improves task performance by 15–30%, enhances reasoning efficiency by 40%, and achieves a 90% manual interpretability score through the Presenter layer. This work provides a practical solution to break through the application bottleneck of MLLMs.