Full-Sequence Grounding on MotionAtlas-Bench
37.6Overall AccuracyGPT-5.2
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GPT-5.2Sampling=16 frames2026.06 | 37.6 | 36 | 38.8 | 36.6 | 38.5 | 46.7 | |
| Gemini 3 ProSampling=16 frames2026.06 | 36.5 | 36.8 | 33.5 | 38.1 | 38.1 | 44 | |
| Qwen3-VL-32BSampling=16 frames2026.06 | 35.7 | 32.8 | 33.9 | 34.4 | 38.4 | 53.4 | |
| Gemini 2.5 FlashSampling=16 frames2026.06 | 35.4 | 34.7 | 36.3 | 32.3 | 35.5 | 47.3 | |
| Qwen3-VL-8B + MotionAtlas-DataFine-tuning data=MotionAtlas-Data, Sampling=16 frames2026.06 | 34.1 | 33.5 | 33.6 | 33 | 31.4 | 45.2 | |
| Qwen3-VL-235BSampling=16 frames2026.06 | 33.7 | 35 | 33.2 | 31.1 | 34.4 | 38.7 | |
| Gemini 2.5 ProSampling=16 frames2026.06 | 31.9 | 33.4 | 30.5 | 25.5 | 34.1 | 47.3 | |
| Qwen3-VL-4B + MotionAtlas-DataFine-tuning data=MotionAtlas-Data, Sampling=16 frames2026.06 | 30.1 | 28.3 | 30.3 | 29.3 | 30.7 | 37.7 | |
| Molmo2-8B + MotionAtlas-DataFine-tuning data=MotionAtlas-Data, Sampling=16 frames2026.06 | 27.5 | 27.4 | 26.7 | 27.3 | 28.4 | 30.4 | |
| Qwen3-VL-8BSampling=16 frames2026.06 | 26.7 | 27.2 | 24.6 | 26.7 | 26.4 | 34.7 | |
| Molmo2-4B + MotionAtlas-DataFine-tuning data=MotionAtlas-Data, Sampling=16 frames2026.06 | 25.5 | 25.4 | 24.8 | 25.2 | 26.4 | 28.3 | |
| Qwen3-VL-4BSampling=16 frames2026.06 | 21.7 | 21.9 | 22.4 | 16.5 | 22.6 | 31.1 | |
| Molmo2-8BSampling=16 frames2026.06 | 21.5 | 21.5 | 21 | 21.4 | 21.9 | 23.9 | |
| Molmo2-4BSampling=16 frames2026.06 | 15.6 | 15.6 | 15 | 15.3 | 16.1 | 18.1 |