RIFT Checkpoint

Official checkpoint for Keep the Future, Drop the Rollout: RIFT for World Action Models.

Project

• Paper: https://arxiv.org/abs/2608.11521
• Code: https://github.com/ChushanZhang/RIFT
• Dataset: https://huggingface.co/datasets/yuanty/LIBERO-fastwam

Files

rift_step021700.pt
dataset_stats.json
config.yaml

• rift_step021700.pt: released RIFT checkpoint for LIBERO
• dataset_stats.json: normalization statistics used by the checkpoint
• config.yaml: current Hydra model configuration using the rift.* namespace

Keep dataset_stats.json with the checkpoint during evaluation.

Install the standalone RIFT repository with pip install -e .. Compose config.yaml with the repository's task and data configs; it contains interpolations such as ${data.train...} and is not a complete training configuration by itself.

rift_step021700.pt uses PyTorch serialization. Download it only from this repository and load it with weights_only=True.

Download

pip install -U huggingface_hub

hf download PoopBear/RIFT \
  rift_step021700.pt \
  dataset_stats.json \
  config.yaml \
  --local-dir ./checkpoints/rift

Evaluation commands are provided in the RIFT repository.

Citation

@article{zhang2026rift,
  title={Keep the Future, Drop the Rollout: RIFT for World Action Models},
  author={Zhang, Chushan and Tong, Jinguang and Li, Xuesong and Wang, Yikai and Li, Hongdong},
  journal={arXiv preprint arXiv:2608.11521},
  year={2026}
}

License

The checkpoint is distributed under the Apache License 2.0. RIFT source code is distributed separately under the MIT License. See NOTICE for attribution.