Video Question Answering on DoraVQA (test)
76.1Overall AccuracyGemini-3.0-Flash
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Gemini-3.0-FlashProprietary=true2026.01 | 76.1 | 67.83 | 81.25 | 65.38 | 88.52 | |
| Gemini-3.0-FlashProprietary=true, Inference protocol=low thinking2026.01 | 75.84 | 66.81 | 87.5 | 69.23 | 86.89 | |
| Qwen3-VL-8B + GRPOModel backbone=Qwen3-VL-8B, Training strategy=GRPO2026.01 | 67.98 | 62.5 | 37.5 | 59.62 | 72.13 | |
| GPT-4VProprietary=true2026.01 | 67.79 | 60.34 | 50 | 69.23 | 68.85 | |
| Gemini-2.5-ProProprietary=true2026.01 | 64.41 | 62.62 | 50 | 60.78 | 71.93 | |
| Qwen2-VL-7B + GRPOModel backbone=Qwen2-VL-7B, Training strategy=GRPO2026.01 | 62.38 | 59.74 | 43.75 | 64.58 | 63.93 | |
| Qwen3-VL-8BModel backbone=Qwen3-VL-8B, Training strategy=baseline2026.01 | 58.08 | 52.16 | 43.75 | 41.18 | 65.57 | |
| InternVideo2.5-8BModel backbone=InternVideo2.5-8B2026.01 | 57.68 | 46.55 | 31.25 | 59.62 | 68.85 | |
| Qwen2-VL-7BModel backbone=Qwen2-VL-7B, Training strategy=baseline2026.01 | 56.74 | 53.02 | 43.75 | 51.92 | 60.66 | |
| LLaVA-Video-7BModel backbone=LLaVA-Video-7B2026.01 | 55.41 | 54.02 | 31.25 | 66 | 67.31 | |
| Qwen2-VL-2B + GRPOModel backbone=Qwen2-VL-2B, Training strategy=GRPO2026.01 | 55.11 | 55.45 | 28.57 | 50 | 50.67 | |
| Qwen2-VL-2BModel backbone=Qwen2-VL-2B, Training strategy=baseline2026.01 | 41.36 | 38.77 | 37.5 | 38.46 | 45.9 | |
| Video-LLaVA-7BModel backbone=Video-LLaVA-7B2026.01 | 37.82 | 47.49 | 41.67 | 46.34 | 32.43 |