HiDrop: Hierarchical Vision Token Reduction in MLLMs
via Late Injection, Concave Pyramid Pruning, and Early Exit

HiDrop: Hierarchical Vision Token Reduction in MLLMs via Late Injection, Concave Pyramid Pruning, and Early Exit

Hao Wu^*,1,2, Yingqi Fan^*,1, Jinyang Dai³, Junlong Tong^1,2,4, Yunpu Ma⁵, Xiaoyu Shen^†,1,2

¹Institute of Digital Twin, Eastern Institute of Technology, Ningbo

²Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative

³University of Science and Technology of China ⁴Shanghai Jiao Tong University

⁵Munich Center of Machine Learning, LMU Munich

^* Equal Contribution, ^† Corresponding Author.

Contact: haowu.ai.research@gmail.com, xyshen@eitech.edu.cn

If you find this work useful for your research and applications, please consider citing:

@misc{wu2026hidrophierarchicalvisiontoken,
      title={HiDrop: Hierarchical Vision Token Reduction in MLLMs via Late Injection, Concave Pyramid Pruning, and Early Exit}, 
      author={Hao Wu and Yingqi Fan and Jinyang Dai and Junlong Tong and Yunpu Ma and Xiaoyu Shen},
      year={2026},
      eprint={2602.23699},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.23699}, 
}

🔥News

• [TODO] Code, checkpoints, and documentation are being prepared and will be released soon.
• [2026.02.27] The preprint is now published!

💡 Highlights (TODO)

📚 Contents

• News: Latest updates, news, and announcements.
• Highlights: Core insights and key features highlighted in this work.
• License: License information for this repository.
• Acknowledgments: Credits to projects and contributors that inspired or supported this work.
• Contact: Contact information for questions, feedback, or collaboration.
• Related Projects: Research projects from our group (EIT-NLP) related to MLLM compression.

📄 License

This project is released under the Apache 2.0 license.

🙏 Acknowledgments

• Thanks for the LLaVA, FastV, and PyramidDrop library, which helps us to quickly implement our ideas.

✉️ Contact

For questions, suggestions, or collaboration opportunities, please feel free to reach out:

• Hao Wu: haowu.ai.research@gmail.com
• Xiaoyu Shen: xyshen@eitech.edu.cn

🌐 Related Projects (ours)

• Survey
  • [Preprint] From Data to Model: A Survey of the Compression Lifecycle in MLLMs
• Vision Encoder
  • [CVPR 26] UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking
• ImageLLM
  • [EMNLP 25] VisiPruner: Decoding Discontinuous Cross-Modal Dynamics for Efficient Multimodal LLMs
  • [Preprint] ViCA: Efficient Multimodal LLMs with Vision-Only Cross-Attention