HappyWorld-Bench

HappyWorld-Bench. By Zhiqi Bai and thirty-five others, from Alibaba Group, the Beijing Institute for General Artificial Intelligence, and three other institutions. Published on the twenty-first of September. No code repository is linked to it.

A world model has to be judged on more than how good its generated world looks. It also has to stay consistent and responsive while something explores it, interacts with it, or modifies it. HappyWorld-Bench is built to evaluate that second property: whether generated worlds remain reliable as agents interact with them. Quality at a glance and reliability under use are different questions, and the benchmark is aimed at the second.

The design rests on a hierarchical capability framework of six world capabilities, labelled W1 through W6. They run from generative construction at one end to unified world modeling at the other. The same framework is instantiated across three independent evaluation tracks: one for video world models, one for spatial world models, and one for embodied world models. Behavioral correctness is captured by newly designed automated metrics, and alongside them sit human judgments. Splitting by track matters because the three kinds of model are asked to do different things. A video model extends footage, a spatial model builds and edits a scene, and an embodied model has to respond to actions.

The test material is sized per track. There are one thousand one hundred thirty- eight video prompts. There are three hundred spatial scenes. There are two hundred fifty-four embodied test cases. Across all three tracks, the authors build and operate HappyWorld-Arena, which organizes human A/B comparisons and derives Elo ratings for each model, complementing the automated metrics. Under this unified framework they evaluate fourteen video world models, nine spatial systems, and eight embodied candidates.

Every track shows reliability gaps that remain. Video models lose consistency during extended rollouts and when a scene is revisited. The best spatial models reach seventy point one four percent placement accuracy. Their best edit execution is seventy-three point three three percent. Embodied models struggle to preserve state across multi-step actions, and to respond precisely when action conditions or physical rules are altered.

The benchmark's argument is about what should count. Visual quality alone does not tell you whether a world model holds together. State consistency, and whether the model responds correctly to actions and interventions, need to be measured too, and on all three tracks those are where current models still fall short. A world that looks right on first viewing can still forget what was there a moment ago, or ignore what an agent just did to it, and those are the failures the benchmark is designed to surface.