HappyWorld-Bench tests reliability of AI generated worlds under interaction
HappyWorld-Bench
Computer Vision and Pattern Recognition
Summary
Creating computer-generated worlds that behave realistically is tricky, especially when these worlds change as people or AI explore and interact with them. The authors created HappyWorld-Bench, a big new test for AI that checks if generated worlds stay consistent and respond correctly as agents interact with them. Their tests cover video sequences, 3D spaces, and embodied agents, revealing that current models still often fail to keep track of changes or behave reliably over time. This work shows that just looking good isn't enough—world models also need to act and update correctly as things happen inside them.
What this means in practice
- •For game developers: Assess AI-generated game environments to ensure reliable interactions and consistent world behaviors under player actions.
- •For robotics engineers: Evaluate embodied AI systems for maintaining accurate state and response during multi-step physical tasks and environmental changes.
Authors
Zhiqi Bai, Junai Cai, Yixin Chen, Jingrun Du, Tao Feng, Wei Gong, Siyuan Huang, Xiao Lin, Jiaheng Liu, Jun Luo, Yongzhe Lyu, Liya Ma, Zenan Meng, Lin Qu, Wenbo Su, Jiaming Wang, Qinghe Wang, Shaofei Wang, Yanghai Wang, Zequn Wang, Ziming Wang, Hu Wei, Jiangtao Wu, Ruiqi Wu, Jiaxin Xie, Yuchi Xu, Ze Xu, Chengting Yu, Liangyu Yuan, Gang Zeng, Yawen Zeng, Xingyao Zhang, Zizheng Zhang, Bo Zheng, Jiancheng Zhu, Song-Chun Zhu
Abstract
Evaluating world models requires assessing both the quality of the worlds they generate and their consistency and responsiveness under exploration, interaction, and modification. We introduce HappyWorld-Bench, a comprehensive benchmark that evaluates whether generated worlds remain reliable as agents interact with them. Our design is built on a hierarchical capability framework of six world capabilities (W1-W6), from generative construction to unified world modeling, instantiated across three independent evaluation tracks: video world models, spatial world models, and embodied world models. HappyWorld-Bench comprises 1,138 video prompts, 300 spatial scenes, and 254 embodied test cases. Across all three tracks, we build and operate HappyWorld-Arena to organize human A/B comparisons and derive model-level Elo ratings, which complement newly designed automated metrics that capture behavioral correctness. We evaluate 14 video world models, 9 spatial systems, and 8 embodied candidates under this unified framework. Results reveal remaining reliability gaps across all three tracks: video models exhibit reduced consistency during extended rollouts and revisits, spatial models achieve at best 70.14% placement accuracy and 73.33% edit execution, and embodied models struggle to preserve state across multi-step actions and respond precisely to altered action conditions and physical rules. These findings highlight the need to evaluate world models not only by visual quality, but also by state consistency and the correctness of their responses to actions and interventions.