Document Question Answering on DocBench
85.5LasJ ScoreDocDancer
Evaluation results
| Method | Links | |
|---|---|---|
| DocDancerModel=GPT-5.22026.01 | 85.5 | |
| MARDocBackbone=Qwen3-VL-32B2026.06 | 83.2 | |
| MARDocBackbone=Qwen3-VL-30B-A3B2026.06 | 82.1 | |
| DocDancerModel=Qwen3-30B-A3B (ft)2026.01 | 81.2 | |
| Human Baseline2026.01 | 81.2 | |
| DocAgentModel=GPT-4o2026.01 | 79.9 | |
| DocDancerModel=Gemini-2.5-Pro2026.01 | 79.9 | |
| DocDancerModel=Qwen3-4B (ft)2026.01 | 79.8 | |
| DocDancerModel=GPT-4o2026.01 | 73.5 | |
| fitzModel=GPT-42026.01 | 67.9 | |
| RAGAnythingModel=GPT-4o-mini2026.01 | 63.4 | |
| Naive VLModel=GPT-4o2026.01 | 63.1 |