The field of document understanding is witnessing a significant shift towards more efficient and scalable solutions, particularly in the context of multimodal large language models (MLLMs). Recent advancements focus on reducing the computational burden associated with handling diverse document formats and sizes, while also enhancing the models' ability to interpret and generate accurate responses. Innovations in hierarchical feature aggregation and instruction tuning are enabling more effective OCR-free document understanding, addressing the limitations of traditional methods. Additionally, the introduction of benchmarks and datasets tailored for multimodal long document understanding and historical manuscript analysis is paving the way for more nuanced and humanities-aligned research. The field is also grappling with the challenges of multimodal knowledge conflicts, proposing fine-tuning methods to ensure consistency between perception and cognition. Furthermore, the integration of visual dialogues into end-to-end retrieval systems is streamlining multimodal query representation, improving retrieval performance. Lastly, there is a growing emphasis on leveraging unstructured text and digitizing local language archives to enhance open-domain dialogue systems and expand NLP research to underrepresented languages.