Single-Frame Grounding on MotionAtlas-Bench
36.9Overall AccuracyGPT-5.2
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GPT-5.2Sampling=16 frames2026.06 | 36.9 | 36.5 | 34 | 34.2 | 42.1 | 51.3 | |
| Gemini 3 ProSampling=16 frames2026.06 | 36.4 | 36.6 | 34.7 | 32 | 43.2 | 46 | |
| Qwen3-VL-8B + MotionAtlas-DataFine-tuning data=MotionAtlas-Data, Sampling=16 frames2026.06 | 31.6 | 31.1 | 31.2 | 30.6 | 29 | 42.8 | |
| Qwen3-VL-235BSampling=16 frames2026.06 | 30.5 | 31.8 | 27.8 | 28.9 | 31.5 | 38.7 | |
| Gemini 2.5 ProSampling=16 frames2026.06 | 29.8 | 29.8 | 28.9 | 27.5 | 29.3 | 43.3 | |
| Qwen3-VL-32BSampling=16 frames2026.06 | 29.7 | 30.3 | 27.5 | 26 | 31.5 | 41.6 | |
| Gemini 2.5 FlashSampling=16 frames2026.06 | 28.9 | 29 | 27.1 | 25.3 | 33 | 40 | |
| Qwen3-VL-4B + MotionAtlas-DataFine-tuning data=MotionAtlas-Data, Sampling=16 frames2026.06 | 27.7 | 25.9 | 27.9 | 26.9 | 28.3 | 35.3 | |
| Molmo2-8B + MotionAtlas-DataFine-tuning data=MotionAtlas-Data, Sampling=16 frames2026.06 | 24.4 | 24.2 | 23.4 | 24.3 | 25.4 | 27.4 | |
| Qwen3-VL-8BSampling=16 frames2026.06 | 24.3 | 24.5 | 23.9 | 20.3 | 24.8 | 37.6 | |
| Molmo2-4B + MotionAtlas-DataFine-tuning data=MotionAtlas-Data, Sampling=16 frames2026.06 | 21.6 | 21.6 | 20.9 | 21.2 | 22.7 | 24.4 | |
| Qwen3-VL-4BSampling=16 frames2026.06 | 19.3 | 19.5 | 20 | 14.1 | 20.2 | 28.7 | |
| Molmo2-8BSampling=16 frames2026.06 | 19.2 | 19.1 | 18.3 | 19.3 | 20 | 21.6 | |
| Molmo2-4BSampling=16 frames2026.06 | 13.7 | 13.6 | 12.9 | 13.7 | 14.5 | 16.1 |