MODEL
Genie3
Google DeepMind
PLAYER-GUIDED EVALUATION FOR INTERACTIVE WORLD MODELS
PlayWorld
Benchmarking World Models with Agent Players over Long-Horizon Objectives
Kuaishou Technology · The University of Hong Kong
171Cases
1,417Videos
10–60sRollouts
9+Models
01 / CURATED MODEL ROLLOUTS
Model demos.
Geometry Consistency60.1s
02 / FULL PLAYER PROCESS
Full Player process.
Complete recordings from world creation to adaptive action execution and final result preservation.
MODEL
HappyOyster
Alibaba
MODEL
HY-World2
Tencent
03 / PIPELINE
Pipeline.
NO HUMAN OPERATIONFULLY AUTOMATED EVALUATION
The Engine enters the initial-world condition, waits for the generated world to become interactive, dispatches controls and captures observations for the Player’s online decisions, then terminates the episode while preserving the final video and execution record.
Comparison of world-model evaluation benchmarks
✓ full coverage · ✕ no coverage · ~ partial coverage
| Benchmark | Input | Closed-loop adaptation | Long-horizon revisit | Geometry consistency | State evolution | Interaction fidelity | Time scale / video |
|---|---|---|---|---|---|---|---|
| WorldScore | Text + Image + Camera | ✕ | ✕ | ✓ | ✕ | ✕ | ~2–10 s |
| WorldMark | Image + Action | ✕ | ~ | ✓ | ✕ | ✕ | 20 / 40 / 60 s |
| WBench | Text + Image + Camera / Action | ✕ | ~ | ✓ | ~ | ✓ | ~5–10 s |
| WorldRoamBench | Image + Action | ✕ | ~ | ✓ | ~ | ✓ | 10–60 s |
| MemoBench | Text + Image + Camera | ✕ | ✕ | ✓ | ✓ | ✕ | ~5–10 s |
| Omni-WorldBench | Text + Image + Camera | ✕ | ✕ | ~ | ✓ | ✓ | 3–6 s |
| PlayWorld (Ours) | Text + Image + Objective | ✓ | ✓ | ✓ | ✓ | ✓ | 10–60 s |
04 / LEADERBOARD
Model ranking.
Select one capability to view its ranking.
| Rank | Model | Score |
|---|---|---|
| 1 | Genie3 | 2.12 |
| 2 | HappyOyster | 1.92 |
| 3 | LingBot-World2 | 1.82 |
| 4 | LingBot-World | 1.78 |
| 5 | HY-World2 | 1.61 |
| 6 | SANA-WM | 1.48 |
| 7 | Hunyuan-GameCraft-2 | 1.42 |
| 8 | HY-WorldPlay | 1.21 |
| 9 | Matrix-Game-3.0 | 1.14 |
Complete nine-model ranking from the current paper results. Equal scores share a rank.
VQA evaluation across four dimensions in PlayWorld
Scores range from 1 to 5. Overall is the unweighted mean across the four evaluation dimensions.
| Model | Geometry consistency ↑ | Interaction fidelity ↑ | Insight evolution ↑ | Out-of-sight evolution ↑ | Overall ↑ |
|---|---|---|---|---|---|
| Closed-source Models | |||||
| Genie3 | 2.74 | 2.40 | 1.51 | 1.81 | 2.12 |
| HappyOyster | 2.54 | 2.15 | 1.47 | 1.54 | 1.92 |
| Open-source Models | |||||
| LingBot-World | 2.11 | 2.23 | 1.33 | 1.43 | 1.78 |
| LingBot-World2 | 2.04 | 2.13 | 1.95 | 1.16 | 1.82 |
| HY-World2 | 2.14 | 2.06 | 1.13 | 1.09 | 1.61 |
| SANA-WM | 1.72 | 1.89 | 1.13 | 1.16 | 1.48 |
| Hunyuan-GameCraft-2 | 1.62 | 1.52 | 1.21 | 1.31 | 1.42 |
| HY-WorldPlay | 1.12 | 1.63 | 1.01 | 1.08 | 1.21 |
| Matrix-Game-3.0 | 1.30 | 1.25 | 1.00 | 1.00 | 1.14 |