HappyWorld-Bench
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
cs.CV
2026-09-21
HappyWorld-Bench scores video, spatial, and embodied world models on a shared W1–W6 ladder. Visual quality is high; best spatial placement is 70.14% and best edit execution 73.33%.
A world model is useful only if its predictions survive interaction. Pretty frames are not enough. After the camera moves, after an object is occluded and revisited, after a rule is edited, the world still has to be the same world.
Existing benchmarks split this requirement. Video suites score appearance and short-horizon control. Spatial suites score scene looks. Embodied suites score action-conditioned video. HappyWorld-Bench puts the three families on one W1–W6 capability ladder and adds HappyWorld-Arena, a human A/B Elo.
Six levels: W1 perceptual construction, W2 action-conditioned local transitions, W3 persistence under long interaction and revisits, W4 programmable interventions whose side effects stay local, W5 unbounded shared worlds, W6 a unified generator-simulator-planner (mostly unscored).
Three tracks share the taxonomy with different interfaces.
All three tracks run a human arena. The embodied arena reports 62,592 blind pairwise votes.
Video, 14 models. HappyOyster leads human Elo at 1263, Genie 3 at 1206. On W1, Genie 3 is slightly ahead (82.8 vs 82.4). On W2/W3 HappyOyster leads (78.0 and 76.8). W4 is currently scored only for two systems that expose interventions: HappyOyster 66.7, LingBot-World-v2 59.3. Post-edit state fidelity is much weaker: 38.61 vs 28.58. Rule edits often change the picture without changing later dynamics.
Spatial, 9 systems. Marble has the highest Elo (1308) and a strong W1 (79.10). Usable geometry is a different ranking: HYWorld-2.0 has the best placement rate at 70.14% and the best edit execution at 73.33%. FlashWorld and both Lyra variants are barely navigable (W2 0.00–1.36). Traversal and support also diverge: WorldGen is near the top in navigable area but places plates only 51.39% of the time.
Embodied, 8 video models. MiniMax-H3 leads all three overall scores (W2 93.60, W3 91.42, W4 88.62) and Arena Elo (1060). Grok Imagine Video 1.5 is close on Elo (1051) with W2 93.02. Sora 2 is last (Elo 872, W2 60.91). Appearance is generally fine. Failures cluster on missed contact, grasping the wrong object, unprompted viewpoint jumps, and W4 material or gravity edits that leave dynamics unchanged.
| Track | Strongest reported signal | Number |
| Video human preference | HappyOyster Elo | 1263 (Genie 3: 1206) |
| Spatial placement | HYWorld-2.0 | 70.14% |
| Spatial edit execution | HYWorld-2.0 | 73.33% |
| Embodied W2 overall | MiniMax-H3 | 93.60 (Sora 2: 60.91) |
The paper is a scoreboard that moves world-model evaluation from generation quality to behavioral correctness. Video groups cannot treat Arena or W1 as the whole story: HappyOyster wins preference and W2/W3, Genie 3 wins W1, and W4 state fidelity falls below 40. Spatial groups cannot treat Marble's looks as navigability; FlashWorld-style systems that render but do not support motion sit near zero on W2. Robot-video groups already produce convincing egocentric clips; the contact chain and condition-dependent physics still fail.
If a world model is meant to be a planner or a data engine, this table says current systems are closer to high-resolution cutscenes than to queryable state machines.
W6 is essentially untested; W5 is only lightly covered via spatial expansion and multi-subject video. The spatial track uses static scenes because no evaluated method does spatiotemporal worlds. The embodied track treats video generators as world models, with language actions rather than joint commands, so it sits a layer above real closed-loop control. Video W4 includes only two models. Many automatic metrics depend on VLMs and Gemini checklists, and near ties can reorder versus human Elo. Evaluation cost is not reported, and no downstream policy success is attached. This is a diagnostic benchmark, not a method that raises a training score.