The World Model Bench (WM Bench) is the world’s first benchmark for evaluating the cognitive capabilities of world models and embodied AI systems. It aims to go beyond traditional image and video quality assessments and focus on the cognitive capabilities of models. This dataset is built around the assessment of world model capabilities, covering three core dimensions: perception, cognition, and embodiment. It is further subdivided into 10 types of tasks, including environmental understanding, entity recognition and classification, and prediction-based reasoning. It also includes 100 diverse scenarios designed to systematically evaluate the model's cognitive and decision-making capabilities in complex environments.

Data fields:

• id: Unique identifier for the sample
• cat: Task category label
• scene_context: Scene context input
• PREDICT: Predictive output, indicating hazard and safety directions.
• MOTION: Action output, describing the emotion of the action.