WorldExam is a hierarchical diagnostic benchmark for controllable video world models. Beyond visual appearance and explicit instruction following, it asks whether a model preserves a coherent world and exhibits inherent reactivity. Across camera-, action-, and language-driven interfaces, WorldExam organizes 1,474 test cases into four diagnostic levels and eight dedicated tasks under a unified evaluation protocol.
Inherent Reactivity: the ability to infer from the scene state how the world should react and to generate plausible consequences not explicitly described in the input
WorldExam’s central diagnostic capability
Visual Quality measures the video's apparent appearance, including perceptual plausibility, temporal stability, and aesthetic quality.
Control Adherence measures whether the controlled camera or subject follows the input control.
Spatial Consistency measures whether the model preserves a coherent world when the camera revisits a previously observed viewpoint.
World Reactivity level evaluates scene-conditioned reactions and goal-directed behaviors beyond what is explicitly specified in the input.
Four diagnostic levels map to eight evaluation tasks, which are assigned to static-scene or dynamic-interaction tracks according to scene assumptions and model applicability. Each track reports task-specific and general metrics.
WorldExam represents controllable behavior as an ordered composition of atomic control units and adapts this control intent into an SE(3) camera trajectory, a discrete action sequence, or a natural-language prompt.
By contrast, the World Reactivity level evaluates scene-conditioned reactions and goal-directed behaviors beyond what is explicitly specified in the input.
The dynamic-interaction track contains Subject Control and the five World Reactivity tasks and applies only to compatible action- and language-driven models; Goal Completion is language-only.
Comparison with representative world-model benchmarks. The table compares supported model paradigms, viewpoints, task coverage, case counts, and evaluated models.
| Benchmark | Model Paradigm |
Viewpoint | WorldExam Evaluation Tasks | #Cases | #Models | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| First Person |
Third Person |
Camera Control |
Subject Control |
Scene Revisit |
Terrain Inter. |
Object Inter. |
Social Inter. |
Physical React. |
Goal Compl. |
||||
| WorldScore | C / L | ✓ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | 3,000 | 20 |
| MIND | A | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | 250 | 2 |
| Omni-WorldBench | C / L | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓† | ✗ | ✓† | ✗ | 1,068 | 18 |
| WorldMark | C / A / L | ✓ | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | 500 | 6 |
| iWorld-Bench | C / A / L | ✓ | ✗ | ✓ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | 4,900 | 14 |
| WBench | C / A / L | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓† | ✗ | ✓† | ✗ | 289 | 20 |
| WorldOlympiad | A / L | ✓ | ✗ | ✓ | ✗ | ✗ | ✗ | ✓† | ✗ | ✓† | ✗ | 1,000 | 8 |
| WorldRoamBench | A | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ | ✗ | 600 | 10 |
| WorldExam (Ours) | C / A / L | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | 1,474 | 20 |
Select an evaluation track and model paradigm. Click column headers to sort the results. All retained metrics are normalized so that higher values indicate better performance.
The task scores diagnose Control Adherence and Spatial Consistency, whereas the general metrics characterize Visual Quality.
| # | Model | Average | Task-Specific Metrics | General Metrics | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Overall ↕ | Task ↕ | General ↕ |
Camera Control ↕ |
Scene Revisit ↕ |
3D Consistency ↕ |
Photometric Consistency ↕ |
Temporal Flickering ↕ |
Aesthetic Quality ↕ |
Imaging Quality ↕ |
||
| 1 | NeoVerse | 85.39 | 93.29 | 77.49 | 97.33 | 89.25 | 98.19 | 70.83 | 93.53 | 52.72 | 72.19 |
| 2 | WorldPlay | 81.61 | 82.63 | 80.58 | 92.74 | 72.51 | 98.69 | 81.06 | 95.90 | 53.92 | 73.31 |
| 3 | InSpatio-World (1.3B) | 81.40 | 85.92 | 76.88 | 85.94 | 85.90 | 98.44 | 64.88 | 93.68 | 53.78 | 73.60 |
| 4 | TrajectoryCrafter | 78.00 | 81.68 | 74.31 | 80.32 | 83.03 | 96.98 | 62.83 | 92.56 | 52.17 | 67.03 |
| 5 | Infinite-World | 74.33 | 64.52 | 84.13 | 71.70 | 57.34 | 99.91 | 92.09 | 95.99 | 56.27 | 76.41 |
| 6 | LingBot-World | 70.39 | 58.27 | 82.50 | 58.19 | 58.35 | 99.59 | 81.52 | 96.53 | 59.87 | 75.00 |
| 7 | Matrix-Game 3.0 | 69.28 | 70.30 | 68.26 | 76.36 | 64.25 | 95.59 | 30.14 | 93.40 | 48.80 | 73.39 |
| 8 | Hailuo 2.3 | 68.64 | 55.99 | 81.28 | 63.29 | 48.70 | 99.38 | 79.88 | 95.33 | 56.30 | 75.52 |
| 9 | ReCamMaster | 67.15 | 53.33 | 80.97 | 38.64 | 68.01 | 99.33 | 81.97 | 95.52 | 54.58 | 73.45 |
| 10 | Yume 1.5 | 66.02 | 54.06 | 77.97 | 75.67 | 32.45 | 98.44 | 67.53 | 95.18 | 53.42 | 75.28 |
| 11 | Voyager | 65.84 | 66.23 | 65.46 | 56.19 | 76.27 | 89.54 | 27.43 | 93.04 | 53.59 | 63.69 |
| 12 | Wan 2.6 I2V | 65.71 | 52.52 | 78.90 | 57.72 | 47.32 | 99.49 | 69.42 | 94.13 | 53.93 | 77.53 |
| 13 | HappyHorse 1.0 | 65.32 | 50.37 | 80.27 | 58.29 | 42.45 | 99.62 | 74.17 | 95.02 | 55.56 | 77.00 |
| 14 | Kling 2.5 | 63.98 | 44.50 | 83.46 | 50.18 | 38.81 | 99.87 | 87.87 | 97.56 | 56.28 | 75.70 |
| 15 | Veo 3.1 | 60.99 | 41.94 | 80.03 | 40.83 | 43.05 | 99.06 | 73.61 | 95.43 | 55.49 | 76.57 |
| 16 | Vidu Q3 | 60.24 | 40.90 | 79.57 | 42.75 | 39.05 | 99.43 | 71.28 | 95.01 | 55.38 | 76.73 |
| 17 | Seedance 1.5 | 59.91 | 44.21 | 75.61 | 49.18 | 39.23 | 97.31 | 56.60 | 94.81 | 54.45 | 74.88 |
| 18 | FantasyWorld | 58.68 | 37.12 | 80.23 | 18.46 | 55.79 | 98.57 | 75.60 | 95.91 | 57.11 | 73.98 |
| 19 | Hunyuan-GameCraft | 57.59 | 41.55 | 73.62 | 41.33 | 41.77 | 93.61 | 53.93 | 93.91 | 54.97 | 71.69 |
| 20 | Astra | 57.16 | 35.37 | 78.95 | 32.59 | 38.15 | 96.46 | 78.41 | 96.38 | 51.71 | 71.78 |
The dynamic-interaction track contains Subject Control and the five World Reactivity tasks and applies only to compatible action- and language-driven models; Goal Completion is language-only.
| # | Model | Average | Task-Specific Metrics | General Metrics | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Overall ↕ | Task ↕ | General ↕ |
Subject Control ↕ |
Terrain Interaction ↕ |
Object Interaction ↕ |
Social Interaction ↕ |
Physical Reaction ↕ |
Goal Completion ↕ |
Subject Consistency ↕ |
Motion Smoothness ↕ |
Aesthetic Quality ↕ |
Imaging Quality ↕ |
||
| 1 | Veo 3.1 | 72.77 | 65.02 | 80.52 | 37.28 | 44.71 | 75.96 | 85.10 | 61.76 | 85.30 | 92.07 | 99.12 | 58.22 | 72.65 |
| 2 | Vidu Q3 | 72.35 | 64.18 | 80.51 | 27.67 | 64.39 | 71.59 | 81.91 | 61.23 | 78.26 | 92.78 | 98.76 | 57.24 | 73.25 |
| 3 | Hailuo 2.3 | 72.03 | 63.37 | 80.69 | 36.49 | 61.57 | 67.01 | 72.45 | 63.84 | 78.86 | 93.25 | 99.31 | 57.74 | 72.47 |
| 4 | HappyHorse 1.0 | 70.57 | 60.60 | 80.54 | 33.11 | 56.30 | 65.70 | 76.17 | 47.01 | 85.33 | 92.69 | 98.85 | 57.37 | 73.23 |
| 5 | Wan 2.6 I2V | 66.60 | 52.29 | 80.91 | 29.02 | 49.21 | 44.56 | 66.40 | 48.15 | 76.39 | 94.46 | 98.15 | 56.30 | 74.72 |
| 6 | Seedance 1.5 | 66.50 | 53.36 | 79.64 | 32.51 | 53.83 | 37.91 | 72.09 | 47.60 | 76.24 | 92.14 | 98.87 | 56.72 | 70.84 |
| 7 | LingBot-World | 60.76 | 39.91 | 81.61 | 55.47 | 24.33 | 25.94 | 60.37 | 33.43 | — | 94.98 | 98.89 | 60.86 | 71.69 |
| 8 | Kling 2.5 | 60.45 | 39.85 | 81.04 | 28.40 | 35.95 | 27.70 | 66.80 | 31.99 | 48.25 | 96.00 | 99.48 | 56.86 | 71.83 |
| 9 | WorldPlay | 57.57 | 37.86 | 77.28 | 49.75 | 27.49 | 33.75 | 51.40 | 26.91 | — | 88.06 | 98.09 | 54.19 | 68.76 |
Representative cases across the eight evaluation tasks are shown below.
Representative test cases by evaluation tasks. Hover the strip to pause.
If you find our work useful, please consider citing:
@article{yang2026worldexam,
title = {WorldExam: Benchmarking World Models from Apparent Appearance to Inherent Reactivity},
author = {Yang, Yuxue and Shang, Shuyao and Wang, Jiahe and Zhou, Zitong and Tan, Liang and Zeng, Junhan and Li, Ruizhi and Li, Junyan and Liu, Yu and Yang, Xiao and Li, Yong and Zhu, Jun and Li, Hongsheng and Tan, Tieniu and Fan, Lue and Zhang, Zhaoxiang},
journal = {arXiv preprint arXiv:2608.02603},
year = {2026}
}