Visual Question Answering on GranuVistaVQA
35.74ROUGE-LQwen3-VL-8B† (GranuRAG)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen3-VL-8B† (GranuRAG)Setting=GranuRAG (C), Fine-tuned=true2026.05 | 35.74 | 46.96 | 70.24 | |
| GranuRAGGenerator Backbone=Qwen-VL-Max2026.05 | 32.27 | 52.19 | 79.3 | |
| Qwen-VL-Max (GranuRAG)Setting=GranuRAG (C)2026.05 | 32.01 | 51.6 | 83.9 | |
| Qwen3-VL-8B†Setting=CoT (B), Fine-tuned=true2026.05 | 31.96 | 46.05 | 63.2 | |
| Qwen-VL-MaxSetting=CoT (B)2026.05 | 30.29 | 48.39 | 74.1 | |
| Embedding RetrievalGenerator Backbone=Qwen-VL-Max, Retrieval Strategy=CLIP-based dense retrieval2026.05 | 29.47 | 45.57 | 63.45 | |
| Qwen3-VL-8B†Setting=Baseline (A), Fine-tuned=true2026.05 | 28.61 | 45.38 | 56.9 | |
| VisRAGGenerator Backbone=Qwen-VL-Max2026.05 | 24.06 | 43.35 | 68.06 | |
| BaselineGenerator Backbone=Qwen-VL-Max2026.05 | 23.79 | 40.83 | 56.4 | |
| Qwen-VL-MaxSetting=Baseline (A)2026.05 | 23.34 | 40.53 | 54.7 | |
| GPT-4.1-Mini (GranuRAG)Setting=GranuRAG (C)2026.05 | 22.72 | 44.04 | 75.1 | |
| RAVQA(PreFLMR)Generator Backbone=Qwen-VL-Max2026.05 | 21.27 | 42.6 | 69.24 | |
| GPT-4.1-MiniSetting=CoT (B)2026.05 | 19.55 | 42.79 | 68.9 | |
| GPT-4o (GranuRAG)Setting=GranuRAG (C)2026.05 | 19.16 | 43.43 | 75.4 | |
| Qwen3-VL-8B (GranuRAG)Setting=GranuRAG (C), Fine-tuned=false2026.05 | 18.82 | 46.49 | 65 | |
| GPT-4.1-MiniSetting=Baseline (A)2026.05 | 18.81 | 40.94 | 56.3 | |
| GPT-4oSetting=CoT (B)2026.05 | 18.41 | 42.09 | 63.97 | |
| Claude-3.5-SonnetSetting=CoT (B)2026.05 | 16.39 | 39.54 | 66.7 | |
| Claude-3.5-Sonnet (GranuRAG)Setting=GranuRAG (C)2026.05 | 16.36 | 41.27 | 80.2 | |
| Claude-3.5-SonnetSetting=Baseline (A)2026.05 | 15.11 | 37.26 | 54.5 | |
| GPT-4oSetting=Baseline (A)2026.05 | 12.83 | 39.95 | 52.7 | |
| Qwen3-VL-8BSetting=CoT (B), Fine-tuned=false2026.05 | 12.23 | 44.46 | 61.17 | |
| Qwen3-VL-8BSetting=Baseline (A), Fine-tuned=false2026.05 | 10.88 | 40.04 | 52.9 |