C2-Evo: Co-Evolving Multimodal Data and Model for Self-Improving Reasoning

arXiv 2025

Xiuwei Chen, Wentao Hu, Hanhui Li, Jun Zhou, Zisheng Chen, Meng Cao, Yihan Zeng, Kui Zhang, Yujie Yuan, Jianhua Han, Hang Xu, Xiaodan Liang

We propose propose C2-Evo, an automatic, closed-loop self-improving framework that jointly evolves both training data and model capabilities.

Tabel of Contents

1. Installation
2. Data Preparation
3. Training
4. Inference
5. Citation
6. Acknowledgement

News

• [2025-12-08] We have released our code. 🔦
• [2025-07-29] Our paper is now accessible at https://arxiv.org/pdf/2502.15130!

Installation

Create a conda environment and install the required packages:

Image Classification

conda create -n transmamba-imagenet python=3.9
conda activate transmamba-imagenet

git clone https://github.com/chen-xw/TransMamba-main.git
cd TransMamba-main
pip install -r requirements.txt
pip install -e causal_conv1d>=1.1.0
pip install -e mamba-1p1p1

Data Preparation

• Download ImageNet-1K from the official website.
• Download CIFAR-10/CIFAR-100 from the official website.
• Download LLaVA-Data from the official website.
• Download LLaVA-COT-100K from the official website.

Training

Training and Inference on Image Classification

python  main.py \
  --model-type 'PlainMamba' \
  --model-teacher 'deit_tiny_patch16_224' \
  --distillation-type 'soft' \
  --output_dir 'output/' \
  --epochs 300 \
  --batch-size 256

Training on VQA

bash worker/SFT_LLaVA_Mamba_Distill.sh

Inference

Our evaluation is based on lmms-eval

Citation

If you find our work useful, please consider citing:

@article{chen2025c2evocoevolvingmultimodaldata,
      title={C2-Evo: Co-Evolving Multimodal Data and Model for Self-Improving Reasoning}, 
      author={Xiuwei Chen and Wentao Hu and Hanhui Li and Jun Zhou and Zisheng Chen and Meng Cao and Yihan Zeng and Kui Zhang and Yu-Jie Yuan and Jianhua Han and Hang Xu and Xiaodan Liang},
      year={2025},
      eprint={2507.16518},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.16518}, 
}
}

Acknowledgement

We would like to thank the following works for their code and models:

• Training: LLaVA