A step-by-step python implementation for our paper "Evolutionary System 2 Reasoning: An Empirical Proof".
@article{ERO,
title={Evolutionary System 2 Reasoning: An Empirical Proof},
author={Zeyuan Ma and Wenqi Huang and Guo-Huan Song and Hongshu Guo and Sijie Ma and Zhiguang Cao and Yue-Jiao Gong},
journal={arXiv preprint arXiv:2512.05760},
year={2025}
}
Install dependencies
pip install -r requirements.txt
Check Hardware conditions
To run ERO, we require the system equipped with:
- 64 GB RAM
- 4 * 64 GB GPUs (every GPU we use is a virtual node of H20)
- 150 GB disk capacity
When additional GPUs are used, the RAM and disk capacity should be increased accordingly.
Download Model
Download the model Qwen2.5-vl-7b-instruct from huggingface and place it in the folder "models/base_model".
Download: https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct
Download ARC Task Data (Optional)
We have provided some of the ARC-1 tasks we used in our experiments. If you want to use other tasks, download and put them in the folder "data". Then, modify the variable "es_list" in "main.py" accordingly.
Download: ARC-1: https://github.com/fchollet/ARC-AGI
Run
Run ERO with default settings:
python main.py
You can also indicate evolution settings in ERO:
python main.py --size 1000 --epochs 12 --scale 0.15
See "main.py" for detailed configurable parameters in ERO.
Save data
In the end of the ERO, the results will be saved in the folder "results" automatically and the best model will be saved in "models/best_models".
Evaluate model
A example can be:
python eval.py --state_dict models/best_models/9a4bb226_state_dict.pth --task 9a4bb226
The JSON file for the task will be searched and read from the folder "data". The output and the score will be printed on the console.