| --- |
| license: apache-2.0 |
| language: |
| - en |
| --- |
| |
| <p align="center" width="100%"> |
| <img src="https://raw.githubusercontent.com/dvlab-research/VisionThink/main/files/VisionThink.jpg" alt="Stanford-Alpaca" style="width: 100%; min-width: 300px; display: block; margin: auto;"> |
| </p> |
|
|
|
|
| # VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning |
|
|
|
|
| [](https://arxiv.org/abs/2507.13348) |
| [](https://huggingface.co/papers/2507.13348) |
| [](https://github.com/dvlab-research/VisionThink/blob/main/LICENSE) |
| <a href='https://huggingface.co/collections/Senqiao/visionthink-6878d839fae02a079c9c7bfe'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Data%20Model-Collection-red'></a> |
|
|
|
|
| ## Senqiao/VisionThink-Smart-Train |
|
|
| This is the training dataset used for our Efficient Reasoning VLM on general VQA tasks. |
|
|
|
|
| **VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning [[Paper](https://arxiv.org/abs/2507.13348)]** <br /> |
| [Senqiao Yang](https://scholar.google.com/citations?user=NcJc-RwAAAAJ), |
| [Junyi Li](https://scholar.google.com/citations?hl=zh-CN&user=zQ0P3JAAAAAJ), |
| [Xin Lai](https://scholar.google.com/citations?user=tqNDPA4AAAAJ), |
| [Bei Yu](https://scholar.google.com/citations?user=tGneTm4AAAAJ), |
| [Hengshuang Zhao](https://scholar.google.com/citations?user=4uE10I0AAAAJ), |
| [Jiaya Jia](https://scholar.google.com/citations?user=XPAkzTEAAAAJ)<br /> |
|
|
|
|
| ## Highlights |
| <p align="center" width="80%"> |
| <img src="https://raw.githubusercontent.com/dvlab-research/VisionThink/main/files/Framework.jpg" alt="Stanford-Alpaca" style="width: 80%; min-width: 300px; display: block; margin: auto;"> |
| </p> |
|
|
| 1. Our VisionThink leverages reinforcement learning to **autonomously** learn whether to reduce visual tokens. Compared to traditional efficient VLM approaches, our method achieves significant improvements on **fine-grained** benchmarks, such as those involving OCR-related tasks. |
|
|
| 2. VisionThink improves performance on **General VQA** tasks while reducing visual tokens by **50%**, achieving **102%** of the original model’s performance across nine benchmarks. |
|
|
| 3. VisionThink achieves strong performance and efficiency by simply resizing input images to reduce visual tokens. We hope this inspires further research into **Efficient Reasoning Vision Language Models**. |
|
|
| ## Video |
| <p align="center" width="85%"> |
| <a href="https://www.youtube.com/watch?v=DGjbFbA5mBw" target="_blank"> |
| <img src="https://raw.githubusercontent.com/dvlab-research/VisionThink/main/files/Video.png" alt="Stanford-Alpaca" style="width: 70%; min-width: 300px; display: block; margin: auto;"> |
| </a> |
| </p> |
| |
|
|
|
|
| ## Citation |
|
|
| If you find this project useful in your research, please consider citing: |
|
|
| > This work is highly motivated by our previous effort on efficient VLMs, [**VisionZip**](https://github.com/dvlab-research/VisionZip), which explores token compression for faster inference. |
|
|
| ``` |
| @article{yang2025visionthink, |
| title={VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning}, |
| author={Yang, Senqiao and Li, Junyi and Lai, Xin and Yu, Bei and Zhao, Hengshuang and Jia, Jiaya}, |
| journal={arXiv preprint arXiv:2507.13348}, |
| year={2025} |
| } |
| @article{yang2024visionzip, |
| title={VisionZip: Longer is Better but Not Necessary in Vision Language Models}, |
| author={Yang, Senqiao and Chen, Yukang and Tian, Zhuotao and Wang, Chengyao and Li, Jingyao and Yu, Bei and Jia, Jiaya}, |
| journal={arXiv preprint arXiv:2412.04467}, |
| year={2024} |
| } |
| ``` |
|
|
|
|
|
|
