Physics violation detection on IntPhys 2
96.4Mean AccuracyHuman
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Human2026.06 | 96.4 | — | — | — | — | |
| GEOPHYS Ensemble (OR)Model Type=Ensemble, Protocol=OR2026.06 | 93.3 | 94 | 91 | 95 | 93 | |
| GEOPHYS Ensemble (Majority)Model Type=Ensemble, Protocol=Majority2026.06 | 77.5 | 78 | 75 | 83 | 75 | |
| GEOPHYSBackbone=VOneNet2026.06 | 61.7 | 66 | 58 | 72 | 53 | |
| GEOPHYSBackbone=CORnet-S2026.06 | 61.1 | 58 | 65 | 67 | 56 | |
| GEOPHYSBackbone=DINOv32026.06 | 60.5 | 66 | 58 | 69 | 51 | |
| GEOPHYSBackbone=V-JEPA 22026.06 | 59.5 | 67 | 58 | 60 | 53 | |
| GEOPHYSBackbone=DINOv22026.06 | 58.8 | 62 | 53 | 58 | 61 | |
| V-JEPA 22026.06 | 57.5 | 75 | 59 | 60 | 59 | |
| Gemini 2.52026.06 | 55.6 | 55 | 64 | 64 | 56 | |
| Lin. probeEvaluation Protocol=leave-one-scene-out cross-validation, Backbone=DINOv32026.06 | 55.5 | 51 | 60 | 59 | 53 | |
| GPT-4o2026.06 | 53.8 | 57 | 60 | 60 | 57 | |
| VidMAEv22026.06 | 53.8 | 65 | 55 | 64 | 60 | |
| V-JEPA2026.06 | 53.8 | 57 | 56 | 68 | 52 | |
| Qwen-VL2026.06 | 52.3 | 54 | 57 | 54 | 51 | |
| Gemini 1.52026.06 | 52.3 | 55 | 57 | 56 | 57 | |
| Cosmos2026.06 | 49.4 | 54 | 31 | 52 | 55 |