MMPFN Paper Accepted to CVPR 2026 Our work on Multi-Modal Prior-data Fitted Networks has been accepted for presentation at CVPR 2026. Check back for the camera-ready version and supplementary materials.

MMPFN is an extension of TabPFN, designed to handle multimodal data — combining tabular, image, and text inputs in a unified learning framework. While TabPFN has shown strong performance on purely tabular datasets, it lacks the ability to integrate heterogeneous modalities.

Comprehensive experiments on datasets show that MMPFN outperforms state-of-the-art baselines, efficiently leveraging diverse data types to enhance predictive performance. This demonstrates the potential of extending prior-data fitted networks into the multimodal domain, offering a scalable and effective solution for heterogeneous data learning.

Conda Environment

conda env create -f environment.yaml

Install

python setup.py develop

Place the checkpoint file and dataset in their respective locations, then update the model_path as shown below:

ln -s /path/to/model/params # symlink parameter
ln -s /path/to/data # symlink data

model_path = Path(__file__).parent/ "parameters" / "tabpfn-v2-classifier.ckpt"

To reproduce the experimental results, you can run run_pad_ufes_20_mmpfn.py, which uses Optuna to explore all hyperparameters.

python run_pad_ufes_20_mmpfn.py

To view the results obtained with the optimized parameters, open and execute the notebook file run_pad_ufes_20_mmpfn.ipynb.

This project follows the original TabPFN license policy(Apache 2.0 with additional attribution requirement): here