WRBench/wrbench-results

WRBench Evaluation Results

Model-level WRBench scores for the current public WRBench release. The default model_scores configuration is the rolling 26-model table (11,100 videos). original/wrbench_23model_results.csv is the frozen paper table (9,600 videos).

These are the aggregate scores shown in the WRBench Leaderboard. Scores are in [0, 1]; higher is better.

Version Note

The paper table remains frozen while the default table can add public models. The 2026-07-13 correction classifies Spatia and LiveWorld as TI2V models that use first-frame extraction. It changes input/source metadata only; scores and support counts are unchanged.

What The Scores Mean

• Cam Align.: whether the generated video follows the requested viewpoint.
• Integ.: visual integrity and readability.
• Vis. spatial: spatial consistency while the relevant object is visible.
• Vis. state: action/state consistency while the relevant object is visible.
• Ret. spatial: spatial consistency after an object leaves view and returns.
• Ret. state: state consistency after an object leaves view and returns.
• Reobs. Support: how often the returned-object case is judgeable. This is a coverage statistic, not a quality score.
• Avg: the main WRBench average over the six score columns.

Load

from datasets import load_dataset

scores = load_dataset("WRBench/wrbench-results", "model_scores", split="default")
df = scores.to_pandas().sort_values("Avg", ascending=False)

Links

• Leaderboard: https://huggingface.co/spaces/WRBench/wrbench-leaderboard
• Videos: https://huggingface.co/datasets/WRBench/wrbench-videos
• Prompts: https://huggingface.co/datasets/WRBench/wrbench-natural25
• Human annotations: https://huggingface.co/datasets/WRBench/wrbench-human-annotations
• Collection: https://huggingface.co/collections/WRBench/wrbench-current-world-models-lack-a-persistent-state-core
• Paper: https://arxiv.org/abs/2606.20545
• Code: https://github.com/JinPLu/WRBench

Citation

@misc{lu2026currentworldmodelslack,
      title={Current World Models Lack a Persistent State Core},
      author={Jinpeng Lu and Dexu Zhu and Haoyuan Shi and Linghan Cai and Guo Tang and Yinda Chen and Jie Cao and Duyu Tang and Yi Zhang and Yong Dai and Xiaozhu Ju},
      year={2026},
      eprint={2606.20545},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.20545},
}