README.md

---
license: apache-2.0
datasets:
- Senqiao/VisionThink-Smart-Train
- Senqiao/VisionThink-Smart-Val
base_model:
- Qwen/Qwen2.5-VL-7B-Instruct
---

<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


[![Paper](https://img.shields.io/badge/Paper-Arxiv%20Link-light)](https://arxiv.org/abs/2507.13348)
[![HF](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Discussion-orange)](https://huggingface.co/papers/2507.13348)
[![Code License](https://img.shields.io/badge/Code%20License-Apache_2.0-yellow.svg)](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-Efficient

This model is trained via reinforcement learning using [`Senqiao/VisionThink-Smart-Train`](https://huggingface.co/datasets/Senqiao/VisionThink-Smart-Train), demonstrating enhanced performance and efficiency 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}
}
```