Detecting physically implausible events on IntPhys2
1Permanence (Fixed) Win RateHuman
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Human2026.02 | 1 | 0.9926 | 0.9711 | 0.9044 | 0.9904 | 0.9444 | 0.9615 | 0.9521 | |
| 3DSPA3D Structure=Enabled, DINOv2 Features=Enabled2026.02 | 0.7692 | 0.75 | 0.7308 | 0.7647 | 0.6731 | 0.6912 | 0.7077 | 0.6447 | |
| 3DSPA3D Structure=Enabled, DINOv2 Features=Disabled2026.02 | 0.6538 | 0.6029 | 0.4615 | 0.6618 | 0.5 | 0.5882 | 0.3846 | 0.3936 | |
| Gemini-2.5 Flash2026.02 | 0.6442 | 0.5882 | 0.5962 | 0.6397 | 0.5481 | 0.5515 | 0.5577 | 0.5638 | |
| VideoMAEv2-g2026.02 | 0.6346 | 0.5 | 0.5481 | 0.5369 | 0.6538 | 0.5441 | 0.4808 | 0.5957 | |
| V-JEPA 2-h2026.02 | 0.6346 | 0.6765 | 0.5192 | 0.5638 | 0.5 | 0.5735 | 0.5 | 0.5213 | |
| 3DSPA3D Structure=Disabled, DINOv2 Features=Enabled2026.02 | 0.6154 | 0.7647 | 0.75 | 0.7308 | 0.7885 | 0.7353 | 0.6923 | 0.5957 | |
| GPT-4o2026.02 | 0.5962 | 0.5882 | 0.5865 | 0.5956 | 0.5481 | 0.5735 | 0.5673 | 0.5532 | |
| V-JEPA-h+RoPEPositional Encoding=RoPE2026.02 | 0.5962 | 0.5735 | 0.5577 | 0.5872 | 0.5769 | 0.75 | 0.4615 | 0.5851 | |
| Gemini-1.5 Pro2026.02 | 0.5577 | 0.5588 | 0.5673 | 0.5673 | 0.548 | 0.548 | 0.5673 | 0.5673 | |
| Qwen-VL 2.52026.02 | 0.5385 | 0.5441 | 0.5673 | 0.5368 | 0.5288 | 0.5441 | 0.5096 | 0.5106 | |
| Cosmos-4BParameters=4B2026.02 | 0.5192 | 0.4118 | 0.5096 | 0.4832 | 0.5385 | 0.5 | 0.4808 | 0.5532 | |
| 3DSPA3D Structure=Disabled, DINOv2 Features=Disabled2026.02 | 0.4423 | 0.5 | 0.5481 | 0.5882 | 0.5385 | 0.4853 | 0.4615 | 0.5213 |