Evaluating world models has largely focused on how realistic the generated worlds look. A new paper, HappyWorld-Bench, argues that this is not enough: what matters is whether those worlds remain reliable once agents begin to explore, interact with, and modify them.
The benchmark is organized around a hierarchical capability framework of six world capabilities, labeled W1 through W6, that range from generative construction to unified world modeling. These capabilities are instantiated across three independent evaluation tracks: video world models, spatial world models, and embodied world models.
To support the evaluation, the authors assembled a substantial test suite comprising 1,138 video prompts, 300 spatial scenes, and 254 embodied test cases. Across all three tracks, they built and operated HappyWorld-Arena, a platform for human A/B comparisons that yields model-level Elo ratings, complemented by newly designed automated metrics intended to capture behavioral correctness.
Using this unified framework, the team evaluated 14 video world models, 9 spatial systems, and 8 embodied candidates. The results point to reliability gaps in every track: video models lose consistency during extended rollouts and revisits; spatial models top out at 70.14% placement accuracy and 73.33% edit execution; and embodied models struggle to preserve state across multi-step actions and to respond precisely when action conditions or physical rules are altered.
The authors conclude that world models should be judged not only by visual quality but also by state consistency and the correctness of their responses to actions and interventions.