flowm123/mnist-world

MNIST World Dataset

This repository contains the MNIST World dataset, used for experiments in the paper Flow Equivariant World Models: Memory for Partially Observed Dynamic Environments, accepted at ICML 2026.

MNIST World is a 2D partially observed video world modeling benchmark designed to evaluate how well models can handle smooth, time-parameterized symmetries and unobserved regions that continue to evolve.

Project Resources

• Paper: arXiv:2601.01075
• Project Page: https://flowequivariantworldmodels.github.io/
• GitHub Repository: https://github.com/hlillemark/flowm

Dataset Configurations

The dataset is provided in several configurations to test different aspects of world modeling:

• dynamic_po: Partially observed environments with dynamic elements (the main benchmark for the paper).
• static_po: Partially observed environments where the world is static.
• dynamic_fo: Fully observed environments with dynamic elements.
• dynamic_fo_no_sm: Fully observed environments with dynamic elements but no self-motion.

Each configuration includes train and validation splits.

Citation

If you find this dataset or the FloWM framework useful, please cite:

@misc{lillemark2026flowequivariantworldmodels,
    title={Flow Equivariant World Models: Memory for Partially Observed Dynamic Environments}, 
    author={Hansen Jin Lillemark and Benhao Huang and Fangneng Zhan and Yilun Du and Thomas Anderson Keller},
    year={2026},
    eprint={2601.01075},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={https://arxiv.org/abs/2601.01075}, 
  }