Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models

Official implementation of the paper: Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models.

---

Overview

Joint-Embedding Predictive Architectures (JEPAs) offer an effective framework for learning world models, but without sufficient constraints, their representation spaces can easily collapse. While recent methods like LeWM prevent this by enforcing a strong isotropic Gaussian prior in the high-dimensional ambient space, this can introduce an overly strong structural bias.

Sub-JEPA relaxes this global constraint by applying Gaussian regularization across multiple random subspaces instead of the original embedding space.

Main Result

Sub-JEPA consistently improves over LeWM across four continuous-control environments.

Method	Two-Room	Reacher	PushT	OGB-Cube
LeWM	84.33 ± 4.23	82.67 ± 4.42	84.67 ± 6.53	67.33 ± 5.01
Sub-JEPA	95.00 ± 2.76	84.00 ± 4.00	89.00 ± 5.33	76.33 ± 5.99

---

Repository Layout

Sub-JEPA/
├── le-wm/                 # Upstream LeWM codebase as a git submodule
├── subjepa.py             # MultiSubspaceSIGReg implementation
├── lewm_subjepa.patch     # Patch that integrates Sub-JEPA into LeWM
└── README.md

Installation

1. Clone the repository

git clone --recursive https://github.com/intcomp/Sub-JEPA.git

cd Sub-JEPA

If you forgot --recursive, you can run:

git submodule update --init --recursive

2. Apply the Sub-JEPA patch to LeWM

git -C le-wm apply ../lewm_subjepa.patch

3. Set up the environment + data (same as LeWM)

Follow the upstream LeWM instructions:

• Installation: https://github.com/lucas-maes/le-wm#using-the-code
• Data layout: https://github.com/lucas-maes/le-wm#data

In particular, LeWM (and this repo) expects datasets and checkpoints under $STABLEWM_HOME (defaults to ~/.stable-wm/).

Pretrained Checkpoints

We provide pretrained Sub-JEPA checkpoints on Hugging Face.

You can download all released checkpoints with:

pip install -U huggingface_hub

hf download intcomp/sub-jepa

Train

Training is identical to LeWM and is configured with Hydra. The patch modifies le-wm/train.py and le-wm/config/train/lewm.yaml to use multi-subspace regularization.

PYTHONPATH=. python le-wm/train.py data=tworoom

Make sure to set your WandB entity and project in le-wm/config/train/lewm.yaml, or disable WandB:

PYTHONPATH=. python le-wm/train.py data=tworoom wandb.enabled=false

Sub-JEPA hyperparameters

All Sub-JEPA knobs live under loss.sigreg in le-wm/config/train/lewm.yaml.

Hyperparameter	Description
loss.sigreg.weight	Overall regularization weight. Same role as the LeWM SIGReg weight.
loss.sigreg.kwargs.num_subspaces	Number of subspaces, denoted as $K$ in the paper.
loss.sigreg.kwargs.subspace_dim	Dimension of each subspace $d_s$. If null, uses embed_dim / K and requires divisibility.
loss.sigreg.kwargs.init_mode	Projection initialization mode.
loss.sigreg.theta	Soft orthogonality penalty weight. Only used by trainable projection variants.

Example: To train the model using the default configuration described in our paper, run:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python le-wm/train.py \
  data=tworoom \
  subdir=tworoom/subjepa \
  loss.sigreg.kwargs.init_mode=orthogonal_frozen \
  loss.sigreg.kwargs.num_subspaces=32 \
  trainer.max_epochs=10

Evaluation

Evaluation configs are located under le-wm/config/eval/.

python le-wm/eval.py --config-name=tworoom.yaml policy=tworoom/subjepa

policy must be the checkpoint path relative to $STABLEWM_HOME, without the _object.ckpt suffix.

Example: Full evaluation with our paper's default seeds:

python le-wm/eval.py --config-name=tworoom.yaml policy=tworoom/subjepa seed=42,100,2026,3407,1234,4444 --multirun

Acknowledgements

This codebase is built on top of the official LeWorldModel implementation. We thank the authors of LeWM for releasing their codebase.

Citation

If you find our work useful in your research, please consider citing:

@misc{zhao2026subjepa,
  title        = {Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models},
  author       = {Zhao, Kai and Nie, Dongliang and Lin, Yuchen and Luo, Zhehan and Gu, Yixiao and Fan, Deng-Ping and Zeng, Dan},
  year         = {2026},
  eprint       = {2605.09241},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url          = {https://arxiv.org/abs/2605.09241}
}