Adaptive Token Boundaries: Integrating Human Chunking Mechanisms into Multimodal LLMs