EasyWAM-Hidden-Cosmos25

EasyWAM-Hidden-Cosmos25

This repository hosts released EasyWAM-Hidden checkpoints built on Cosmos-Predict2.5-2B. EasyWAM-Hidden uses a separate Action DiT conditioned one-way on intermediate Video DiT features. Inference retains future-video prediction to provide the hidden visual conditioning. The checkpoints are trained with the EasyWAM codebase.

LIBERO

Results

Task success rate (%) under the EasyWAM LIBERO evaluation protocol. All results use state_position: sequence; higher is better. Models using the same Cosmos-Predict2.5-2B backbone are shown together for comparison, and 🔥 marks the architecture released in this repository.

Full-Parameter

LIBERO-Plus

LIBERO-Plus results are reported for full-parameter checkpoints only. See the complete EasyWAM benchmark table for source results and comparisons across backbones.

Download

hf download OpenMOSS-Team/EasyWAM-Hidden-Cosmos25 \
  libero_hidden_cosmos25.pt \
  libero_dataset_stats.json \
  --local-dir ./checkpoints

Evaluation

Install EasyWAM, prepare the Cosmos-Predict2.5-2B dependencies using the backbone guide, and set up the simulator using the LIBERO evaluation guide. Then run:

# Full-parameter checkpoint
python experiments/libero/run_libero_manager.py \
  task=libero_easywam_hidden_cosmos25 \
  ckpt=./checkpoints/libero_hidden_cosmos25.pt \
  EVALUATION.dataset_stats_path=./checkpoints/libero_dataset_stats.json

The default protocol evaluates all four LIBERO suites with 50 trials per task. Add MULTIRUN.num_gpus=<gpu-count> to distribute evaluation across multiple GPUs.

Available Checkpoints

• LIBERO full-parameter checkpoint: libero_hidden_cosmos25.pt
• LIBERO normalization statistics: libero_dataset_stats.json
• State-token placement: sequence
• Checkpoint format: EasyWAM PyTorch checkpoint (.pt)

Project

• Project page: http://openmoss.ai/EasyWAM/
• Code: https://github.com/OpenMOSS/EasyWAM

License and Citation

EasyWAM code is released under the MIT License. Released checkpoints in this repository use CC BY-NC 4.0 and remain subject to the terms of their base models and training data. If EasyWAM is useful in your research, please cite:

@misc{easywam2026,
  title  = {EasyWAM: A Unified and Efficient Framework for Training and Evaluating World Action Models},
  author = {EasyWAM-Team},
  year   = {2026},
  url    = {https://github.com/OpenMOSS/EasyWAM}
}