JustAct checkpoints

Weights for JustAct: Extracting Action Representations from Foundation Models for World Models (NeurIPS 2026).

• Code, usage and the evaluation of every table: https://github.com/runjiali-rl/justact
• Project page: https://runjiali-rl.github.io/justact/

Download

# in the root of the GitHub repository (README.md is excluded so that checkpoints/README.md is not overwritten)
hf download liguang0115/justact --local-dir checkpoints --exclude README.md .gitattributes
cd checkpoints && sha256sum -c SHA256SUMS

Files

file	size	content	used by
justact/adapter_config.json, justact/adapter_model.safetensors	34 MB	JustAct LoRA adapters for V-JEPA 2 ViT-L (PEFT format)	everything
justact/projection.safetensors	2.6 MB	the fixed random 1024 -> 640 projection (weight, bias)	everything
adaworld_lam_retrained.safetensors	1.9 GB	AdaWorld latent action model trained on our frame pairs	Tables 1, 2, 4
dreamdojo_lam_retrained.safetensors	2.8 GB	DreamDojo latent action model trained on our frame pairs	Tables 1, 2, 4
worldmodel/{no_action,adaworld_lam,dreamdojo_lam,justact}.safetensors	223-246 MB each	trained weights (LoRA + action conditioning) of the four AdaWorld world models	Tables 2, 3
worldmodel/justact_whitening.safetensors	3.3 MB	whitening statistics of the JustAct input of the world model	Tables 2, 3
downstream/idm_{dmlab,minecraft,habitat}.safetensors, downstream/readout_nuscenes.safetensors	about 1 MB each	inverse-dynamics classifiers (Follow metric) and the nuScenes forward-motion readout	Table 3
downstream/step0/*.safetensors, downstream/standardise/*.safetensors	small	action interfaces before fine-tuning and their standardisation statistics	Table 3

Not included, downloaded from their authors: V-JEPA 2 ViT-L (facebook/vjepa2-vitl-fpc64-256, MIT licence), the AdaWorld world model (Little-Podi/AdaWorld, Apache License 2.0) and the OpenCLIP ViT-H-14 image encoder. See checkpoints/README.md in the GitHub repository for where each one goes.

Licence

These weights are released under the Apache License 2.0. The world-model weights are adapters on the AdaWorld world model, which is itself fine-tuned from Stable Video Diffusion, so Stability AI's terms for that model also apply. The training and evaluation datasets keep their own licences, and several of them are for non-commercial use only (listed in the paper's appendix and in the GitHub repository).

Citation

@inproceedings{li2026justact,
  title     = {JustAct: Extracting Action Representations from Foundation Models for World Models},
  author    = {Li, Runjia and Chen, Minghao and Xie, Junyu and Torr, Philip and Vedaldi, Andrea and Jakab, Tomas},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}