Neural Force Field: Few-shot Learning of Generalized Physical Reasoning

ICLR 2026

Shiqian Li^*   Ruihong Shen^*   Yaoyu Tao^†   Chi Zhang^†   Yixin Zhu^†  

Peking University    

^* Equal Contribution     ^† Corresponding Author

📊 Project structure

NeuralForceField/
│
├── data/
│   └──iphyre/game_seq_data/
│       ├── activated_pendulum/
│       ├── angle/
│       ├── ...
│       └── support_hole/
│   └── nbody/
│       ├── train_data.npy
│       ├── val_data.npy
│       ├── within_data.npy
│       └── cross_data.npy
│
├── checkpoints/
│   └── nbody/
│       ├── nff/
│       ├── in/
│       ├── slotformer/
│       └── segno/
│           ├── model_best.pth
│           └── train_args.json
│   └── iphyre/
│
├── iphyre/             
│   ├── configs/         # Configuration files
│   ├── models/          # Model dir containing NFF, IN, SlotFormer
│   ├── utils/           # Useful tools such as dataloader
│   ├── planning.py      # Planning script
│   ├── README.md        # An instruction for use
│   ├── test.py          # Evaluation functions
│   └── train.py         # Training functions
│
├── iphyre/             
│   ├── configs/         # Configuration files
│   ├── models/          # Model dir containing NFF, IN, SlotFormer
│   ├── utils/           # Useful tools such as dataloader
│   ├── generate_data.py # Data generation functions
│   ├── planning.py      # Planning script
│   ├── README.md        # An instruction for use
│   ├── test.py          # Evaluation functions
│   └── train.py         # Training functions

🔧 Getting started

Make sure you have installed torch, torchdiffeq, iphyre, and rebound.

Go to the specific task directory to train and test the models. The instructions of running commands are provided for each task (README_iphyre and README_nbody). Download data here and checkpoin ts here.

cd ./iphyre

or

cd ./nbody

🚀 Related Projects

• [ICLR 2026] Learning Physics-Grounded 4D Dynamics with Neural Gaussian Force Fields: Our latest approach extending NFF to video prediction and 3D deformable objects. [Project Page | Paper | Code]

📚 Citation

If you find our work helpful, please consider citing:

@inproceedings{
  li2026nff,
  title     = {Neural Force Field: Few-shot Learning of Generalized Physical Reasoning},
  author    = {Li, Shiqian and Shen, Ruihong and Tao, Yaoyu and Zhang, Chi and Zhu, Yixin},
  year      = {2026},
  booktitle = {ICLR},
  url       = {https://neuralforcefield.github.io/}
}