license: mit
pipeline_tag: image-text-to-text
library_name: transformers
CodePlot-CoT: Mathematical Visual Reasoning by Thinking with Code-Driven Images
This repository contains the CodePlot-CoT model, a core component of the paper CodePlot-CoT: Mathematical Visual Reasoning by Thinking with Code-Driven Images. CodePlot-CoT is an innovative code-driven Chain-of-Thought (CoT) paradigm designed to enable Vision Language Models (VLMs) to "think with images" when solving mathematical problems. Instead of generating pixel-based images directly, the model outputs executable plotting code to represent its "visual thoughts". This code is then executed to render a precise figure, which is reinput to the model as a visual input for subsequent reasoning steps.
The model is built upon the Qwen2.5-VL architecture and is compatible with the transformers library.
Sample Usage
Installation
To get started with CodePlot-CoT, clone the repository and install the required packages:
conda create -n codeplot python==3.10
conda activate codeplot
git clone [email protected]:HKU-MMLab/Math-VR-CodePlot-CoT.git
cd CodePlot-CoT
pip install -r requirements.txt
pip install flash_attn==2.7.4.post1
For benchmark evaluation only (additional dependencies):
pip install openai==4.1.1
pip install datasets==2.0.0
Model Weights
Ensure your directory structure for the models looks like this:
CodePlot-CoT
├── ckpts
│ ├── CodePlot-CoT
│ ├── MatPlotCode
├── ...
Inference
You can perform inference using the provided scripts:
# Convert image to python code with MatPlotCode
python image_to_code.py
# Solve math problems with CodePlot-CoT
python math_infer.py
For more details on evaluation and benchmarks, please refer to the project homepage and the GitHub repository.
Citation
If you find this work helpful, please consider citing our paper:
@article{duan2025code,
title={CodePlot-CoT: Mathematical Visual Reasoning by Thinking with Code-Driven Images},
author={Duan, Chengqi and Fang, Rongyao and Wang, Yuqing and Wang, Kun and Huang, Linjiang and Zeng, Xingyu and Li, Hongsheng and Liu, Xihui},
journal={arXiv preprint arXiv:2510.11718},
year={2025}
}