WBench: A Comprehensive Multi-turn Benchmark for Interactive Video World Model Evaluation

Researchers introduced WBench, a new benchmark for interactive video world models that uses 289 cases and 1,058 turns.

It evaluates models across five dimensions, including quality, setting adherence, consistency, navigation, and physics compliance, with mixed control interfaces.

Automatic metrics were validated for multi-turn scenarios, including first-person and third-person views and 6-DoF pose control.

Testing 20 models showed no single system is strong across all tasks, with top models excelling in different areas.