Modality Gap-Driven Subspace Alignment Training Paradigm For Multimodal Large Language Models - Explained Simply | ArXiv Explained