Automated Model Discovery
via Multi-modal & Multi-step Pipeline

Project Page | Paper

[NeurIPS'25] Official Repository for 'Automated Model Discovery via Multi-modal & Multi-step Pipeline'

Highlights

Environment Setup

This code was developed at Ubuntu 20.04, using python=3.11, GPy==1.13.2, and GPy-ABCD==1.2.3. Later versions should work, but it have not been tested.

conda create -n amd python=3.11
conda activate amd
pip install -r requirements.txt

Getting Started with Kernel Structure Discovery

Since our model discovery is done in training-free method, you can run the code

Dataset Preparation

Univariate Time-series Dataset are already provided in Automated-Model-Discovery/data, in csv file. They are originated from https://github.com/jamesrobertlloyd/gpss-research.git.

Running the Code

You can start with the code by simply executing:

python3 main_gp.py --noise 0 --data 1 --set_se_const True --model gpt-4o-mini

Getting Started with Symbolic Regression

Dataset Preparation

You can download dataset for Symbolic Regression at https://github.com/merlerm/In-Context-Symbolic-Regression.git/data, and please place them at Automated-Model-Discovery/data/symbolic_regression.

Running the Code

You can start with the code by simply executing:

python3 main_sr.py experiment/function=nguyen/nguyen1

Citation

If you find our code or paper helps, please consider citing our paper:

@inproceedings{
jung-mok2025automated,
title={Automated Model Discovery via Multi-modal \& Multi-step Pipeline},
author={Lee Jung-Mok and Nam Hyeon-Woo and Moon Ye-Bin and Junhyun Nam and Tae-Hyun Oh},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=qGFvTIMS3W}
}

Acknowledgement

We sincerly thank for the authors of these project for making their work publicly available:

• GPy: Gaussian process framework in Python
• GPy-ABCD: A Configurable Automatic Bayesian Covariance Discovery Implementation
• In-Context-Symbolic-Regression: Leveraging Large Language Models for Function Discovery