ZOO-Prune: Training-Free Token Pruning via Zeroth-Order Gradient Estimation in Vision-Language Models

Youngeun Kim, Youjia Jang, Huiling Liu, Aechon Jung, Sunwoo Lee, Sungeun Hong

CVPR, 2026

PDF
arXiv

ZOO-Prune is a training-free VLM token pruning method that uses zeroth-order perturbations at the projection layer to estimate token sensitivity without backpropagation.

BibTeX

@inproceedings{kim2026zoo,
  title={Zoo-prune: Training-free token pruning via zeroth-order gradient estimation in vision-language models},
  author={Kim, Youngeun and Zhang, Youjia and Liu, Huiling and Jung, Aecheon and Lee, Sunwoo and Hong, Sungeun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={39572--39582},
  year={2026}
}