Deictic Visual Question Answering on D-VQA Hybrid (test)
81.1AccuracyHuman Study
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Human Studyagent=human participants2026.03 | 81.1 | 63.1 | 65.3 | |
| LLaVA-OneVision-8Bparameter_count=8B2026.03 | 53.1 | 51.9 | 54.8 | |
| LLaVA-OneVision-4Bparameter_count=4B2026.03 | 51 | 53.5 | 54.7 | |
| Qwen3-VL-8Bparameter_count=8B2026.03 | 49.4 | 48.4 | 49 | |
| Qwen3-VL-4Bparameter_count=4B2026.03 | 47.1 | 48.8 | 49.5 | |
| Qwen2.5-VL-7Bparameter_count=7B2026.03 | 45.7 | 42.7 | 44.5 | |
| Qwen2.5-VL-3Bparameter_count=3B2026.03 | 45.6 | 46.1 | 47 | |
| InternVL-3.5-4Bparameter_count=4B2026.03 | 44.5 | 46.1 | 44.5 | |
| DeepSeek-VL2-Smallvariant=Small2026.03 | 41.3 | 45.9 | 45.5 | |
| DeepSeek-VL22026.03 | 41.2 | 46.1 | 47.3 | |
| GroundingGPT2026.03 | 33.4 | 36.3 | 36.4 | |
| InternVL-3.5-8Bparameter_count=8B2026.03 | 31.6 | 39.2 | 38.1 | |
| LLaVA-v1.5-13Bparameter_count=13B2026.03 | 26.7 | 24.9 | 24.5 |