AVM-Pref

AVM-Pref (Action and Video Mismatch Preference Dataset) provides complementary preference supervision for training reward models that jointly assess action-following fidelity and visual quality in action-conditioned world models. It is introduced in ACWM-RM: A Reference-Free Reward Model from Complementary Mismatch Preferences.

Dataset Overview

The dataset described in the paper contains approximately 506.3k preference pairs, constructed from the AgiBotWorld training subset:

Pair type	Size	Training supervision
Action mismatch	156.3k	Original versus numerically or semantically perturbed action sequences for the same real video, supervising action-following fidelity.
Video mismatch	350k	Generated videos under the same initial frame and action sequence, ranked using global and local PSNR and LPIPS to supervise visual quality. Includes intra-model and cross-model pairs.

Generated videos come from DreamDojo-2B, DreamDojo-14B, GE-Sim-2B, and OSCAR-2B. ACWM-RM learns from these preferences using the Bradley-Terry loss. For construction and training details, please refer to the paper.

Files

train/
  action_mismatch/             # Training action-mismatch pairs (Parquet)
  video_mismatch/              # Training video-mismatch pairs (Parquet)
val/
  action_mismatch/             # Validation action-mismatch pairs (Parquet)
  video_mismatch/              # Validation video-mismatch pairs (Parquet)
_media/
  agibotworld/observations/    # Real-video archives by task
  wm/                         # Generated-video archives by world model and split

For evaluation, see AVM-Eval.

Citation

If you use AVM-Pref in your research, please cite:

@misc{hong2026acwmrm,
  title         = {{ACWM-RM}: A Reference-Free Reward Model from Complementary Mismatch Preferences},
  author        = {Shibo Hong and Siyu Ren and FengJiao Chen and Zhongyuan Peng and Wei Shi and Yixin Cao},
  year          = {2026},
  eprint        = {2610.32222},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.32222}
}