@inproceedings{fan-etal-2025-visipruner,
    title = "{V}isi{P}runer: Decoding Discontinuous Cross-Modal Dynamics for Efficient Multimodal {LLM}s",
    author = "Fan, Yingqi  and
      Zhao, Anhao  and
      Fu, Jinlan  and
      Tong, Junlong  and
      Su, Hui  and
      Pan, Yijie  and
      Zhang, Wei  and
      Shen, Xiaoyu",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-emnlp/2025.emnlp-main.955/",
    pages = "18896--18913",
    ISBN = "979-8-89176-332-6",
    abstract = "Multimodal Large Language Models (MLLMs) have achieved strong performance across vision-language tasks, but suffer from significant computational overhead due to the quadratic growth of attention computations with the number of multimodal tokens. Though efforts have been made to prune tokens in MLLMs, *they lack a fundamental understanding of how MLLMs process and fuse multimodal information*. Through systematic analysis, we uncover a three-stage cross-modal interaction process: (1) Shallow layers recognize task intent, with visual tokens acting as passive attention sinks; (2) Cross-modal fusion occurs abruptly in middle layers, driven by a few critical visual tokens; (3) Deep layers discard vision tokens, focusing solely on linguistic refinement. Based on these findings, we propose *VisiPruner*, a training-free pruning framework that reduces **99.9{\%}** of vision-related attention computations and **62.8{\%}** of FLOPs while maintaining performance. It significantly outperforms existing token pruning methods and generalizes across diverse MLLMs. Beyond pruning, our insights further provide actionable guidelines for training efficient MLLMs by aligning model architecture with its intrinsic layer-wise processing dynamics."
}