MLLM-Microscope: Unlocking Hidden Structure Within Multimodal Large Language Models | ArxivCSExplorer