Deictic Visual Question Answering on D-VQA Real-World (test)
87.7AccuracyHuman Study
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Human Studyagent=human participants2026.03 | 87.7 | 75.6 | 75.3 | |
| LLaVA-OneVision-8Bparameter_count=8B2026.03 | 54.4 | 52.5 | 52.6 | |
| LLaVA-OneVision-4Bparameter_count=4B2026.03 | 52.1 | 47.6 | 48.7 | |
| Qwen3-VL-8Bparameter_count=8B2026.03 | 51.7 | 51 | 50.3 | |
| DeepSeek-VL2-Smallvariant=Small2026.03 | 45.7 | 52.3 | 50.6 | |
| Qwen3-VL-4Bparameter_count=4B2026.03 | 45.6 | 50.7 | 48.9 | |
| Qwen2.5-VL-7Bparameter_count=7B2026.03 | 44.4 | 42.8 | 41.1 | |
| Qwen2.5-VL-3Bparameter_count=3B2026.03 | 44.1 | 52.7 | 52.3 | |
| InternVL-3.5-4Bparameter_count=4B2026.03 | 43.3 | 36.8 | 36.7 | |
| DeepSeek-VL22026.03 | 39.3 | 39.7 | 39 | |
| GroundingGPT2026.03 | 39.2 | 35.2 | 35 | |
| InternVL-3.5-8Bparameter_count=8B2026.03 | 28.6 | 39.8 | 38 | |
| LLaVA-v1.5-13Bparameter_count=13B2026.03 | 14.7 | 15 | 13.7 |