WRBench

WRBench is an open benchmark for camera-controlled video world models and diagnostic D1-D6 evaluation, introduced in Current World Models Lack a Persistent State Core.

Start Here

• Leaderboard: https://huggingface.co/spaces/WRBench/wrbench-leaderboard
• Results dataset: https://huggingface.co/datasets/WRBench/wrbench-results
• Benchmark videos: https://huggingface.co/datasets/WRBench/wrbench-videos
• Natural-25 prompts and first frames: https://huggingface.co/datasets/WRBench/wrbench-natural25
• Human annotation labels: https://huggingface.co/datasets/WRBench/wrbench-human-annotations
• Release collection: https://huggingface.co/collections/WRBench/wrbench-current-world-models-lack-a-persistent-state-core

The paper reports a frozen benchmark table. The Hugging Face results dataset and leaderboard can be refreshed as new public results are released, so current model, video, and support counts should be read from the data files.

Links

• Paper: https://arxiv.org/abs/2606.20545
• Project page: https://jinplu.github.io/WRBench/
• Code: https://github.com/JinPLu/WRBench