[CVPR 2026] TransPrune: Token Transition Pruning for Efficient Large Vision-Language Model

📢 News

• [2026-02-23] 🚀 Source code for TransPrune are now available!
• [2026-02-21] 🎉 Our paper has been accepted by CVPR 2026!
• [2025-07-15] 📄 ArXiv version is released: 2507.20630.

🔧 Install

1. Clone this repository

git clone https://github.com/liaolea/TransPrune.git
cd TransPrune

2. Install Package

conda create -n transprune python=3.10 -y
conda activate transprune
pip install --upgrade pip  # enable PEP 660 support
pip install -e .

📊 Evaluation

We provide scripts to evaluate TransPrune on standard Large Vision-Language Model benchmarks. See .scrpts/transprune/

Cite us

@misc{li2025transprunetokentransitionpruning,
      title={TransPrune: Token Transition Pruning for Efficient Large Vision-Language Model}, 
      author={Ao Li and Yuxiang Duan and Jinghui Zhang and Congbo Ma and Yutong Xie and Gustavo Carneiro and Mohammad Yaqub and Hu Wang},
      year={2025},
      eprint={2507.20630},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.20630}, 
}