---
base_model:
- liuhaotian/llava-v1.5-7b
license: cc-by-nc-nd-4.0
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- multimodal
- chain-of-thought
---

# UV-CoT: Unsupervised Visual Chain-of-Thought Reasoning via Preference Optimization

This repository hosts the **UV-CoT** model, presented in the paper [Unsupervised Visual Chain-of-Thought Reasoning via Preference Optimization](https://huggingface.co/papers/2504.18397).

*   **Project page:** [https://kesenzhao.github.io/my_project/projects/UV-CoT.html](https://kesenzhao.github.io/my_project/projects/UV-CoT.html)
*   **Code:** [https://github.com/UV-CoT](https://github.com/UV-CoT)

## Overview

Chain-of-thought (CoT) reasoning greatly improves the interpretability and problem-solving abilities of multimodal large language models (MLLMs). Existing approaches primarily focus on text CoT, limiting their ability to leverage visual cues. Unsupervised Visual CoT (UV-CoT) introduces a novel framework for image-level CoT reasoning via preference optimization, eliminating the need for extensive labeled bounding-box data.

UV-CoT achieves this by performing preference comparisons between model-generated bounding boxes. It generates preference data automatically, then uses an evaluator MLLM (e.g., OmniLLM-12B) to rank responses, which serves as supervision to train the target MLLM (e.g., LLaVA-1.5-7B). This approach emulates human perception—identifying key regions and reasoning based on them—thereby improving visual comprehension, particularly in spatial reasoning tasks.

<!-- ![Figure 1: UV-CoT Overview](./images/fig1.svg) -->

<div align="center">
  <img src="./images/fig1.svg" alt="Figure 1: UV-CoT Overview" width="1200px" />
</div>

## Visualizations

Qualitative examples demonstrating UV-CoT's visual reasoning:

<div align="center">
  <img src="./images/fig5_v1.2.svg" alt="Figure 5: UV-CoT Visualization 1" width="1200px" />
</div>

<div align="center">
  <img src="./images/fig6_v1.2.svg" alt="Figure 6: UV-CoT Visualization 2" width="1200px" />
</div>

<!-- ![Figure 5: UV-CoT Visualization 1](./images/fig5_v1.2.svg)
![Figure 6: UV-CoT Visualization 2](./images/fig6_v1.2.svg) -->

## Installation

To set up the environment and install necessary packages, follow these steps:

1.  Clone this repository and navigate to the `UV-CoT` folder:
    ```bash
    git clone https://github.com/UV-CoT
    cd UV-CoT
    ```

2.  Create a conda environment and install the package:
    ```bash
    conda create -n uv-cot python=3.10 -y
    conda activate uv-cot
    pip install -e .
    ```
3.  Install the required spaCy model:
    ```bash
    wget https://github.com/explosion/spacy-models/releases/download/en_core_web_trf-3.7.3/en_core_web_trf-3.7.3.tar.gz
    pip install en_core_web_trf-3.7.3.tar.gz
    ```

## Usage

You can load and use the UV-CoT model with the `transformers` library. For detailed information on preference data curation, training, and evaluation, please refer to the [official GitHub repository](https://github.com/UV-CoT).

Here's a basic example of how to use the model for inference:

```python
from transformers import AutoProcessor, AutoModelForCausalLM
from PIL import Image
import requests
import torch

# Load model and processor
model_id = "kesenZhaoNTU/UV-CoT" # Use this model_id to load UV-CoT
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)

# Load an example image
image_url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/bird.jpg"
image = Image.open(requests.get(image_url, stream=True).raw).convert("RGB")

# Define the conversation prompt
prompt = "Describe the image in detail."
messages = [
    {"role": "user", "content": f"<image>
{prompt}"}
]

# Apply the chat template to format the prompt for the model
text = processor.apply_chat_template(messages, add_generation_prompt=True)

# Prepare inputs for the model
inputs = processor(text=text, images=image, return_tensors="pt").to(model.device)

# Generate response
output = model.generate(**inputs, max_new_tokens=200)
print(processor.decode(output[0], skip_special_tokens=True))
```

## Citation

If our work assists your research, feel free to give us a star ⭐ or cite us using:

```bibtex
@misc{zhao2025unsupervisedvisualchainofthoughtreasoning,
      title={Unsupervised Visual Chain-of-Thought Reasoning via Preference Optimization}, 
      author={Kesen Zhao and Beier Zhu and Qianru Sun and Hanwang Zhang},
      year={2025},
      eprint={2504.18397},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.18397}, 
}
```