World Modeling on Manhattan taxi rides
100Next-Token Test AccuracyGPT
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GPTImplementation=nanoGPT [Karpathy, 2022]2025.11 | 100 | 97 | 0.65 | 160.1 | 85 | |
| MTPMulti-step prediction horizon (d)=8, Implementation=Gloeckle et al. [2024]2025.11 | 100 | 98.1 | 0.64 | 57.7 | 95 | |
| JTPMulti-step prediction horizon (d)=8, Reference=Ahn et al. [2025]2025.11 | 100 | 97.1 | 0.32 | 215.8 | 87 | |
| NextLatMulti-step prediction horizon (d)=82025.11 | 100 | 98.7 | 0.71 | 52.7 | 95 | |
| True world model2025.11 | 100 | 100 | 1 | — | 100 |