WBench: A Comprehensive Multi-turn Benchmark for Interactive Video World Model Evaluation
TL;DR AI
2 min readKey summary
Researchers released WBench, a new benchmark for interactive video world models.
It includes 289 cases and 1,058 interaction turns, covering four interaction types and five evaluation dimensions.
The benchmark, code, and data are now available on arXiv and GitHub.
Results across 20 tested models show uneven performance on setting adherence, physics, consistency, navigation, and perspective switching.
WBench provides a shared, broader standard for evaluating these models more rigorously.
