Knowledge-intensive video QA on EncVQA
47.64AccuracyVideoSpeculateRAG
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| VideoSpeculateRAGRAG Strategy=Speculate RAG, Model size=3B+32B2026.01 | 47.64 | 16.61 | 69.38 | 21.18 | |
| Qwen-2.5VL-Instruct-32BRAG Strategy=Standard RAG, Model size=32B2026.01 | 46.75 | 40.42 | 69.03 | 44.07 | |
| InternVL-3-38BRAG Strategy=Standard RAG, Model size=38B2026.01 | 43.5 | 7.75 | 61.18 | 10.88 | |
| VideoSpeculateRAGRAG Strategy=Speculate RAG, Model size=3B+7B2026.01 | 41 | 13.03 | 61.8 | 19.54 | |
| Qwen-2.5VL-Instruct-3BRAG Strategy=Standard RAG, Model size=3B2026.01 | 39 | 4.76 | 57.13 | 4.8 | |
| LLaVA-NeXT-Video-34BRAG Strategy=Standard RAG, Model size=34B2026.01 | 37.5 | 51.37 | 57.15 | 53.96 | |
| Qwen-2.5VL-Instruct-32BRAG Strategy=No RAG, Model size=32B2026.01 | 22 | 20.13 | 39.87 | 19.65 | |
| Qwen-2.5VL-Instruct-3BRAG Strategy=No RAG, Model size=3B2026.01 | 14 | 3.3 | 25.81 | 2.72 | |
| InternVL-3-38BRAG Strategy=No RAG, Model size=38B2026.01 | 13.5 | 5.61 | 26.85 | 4.84 | |
| LLaVA-NeXT-Video-34BRAG Strategy=No RAG, Model size=34B2026.01 | 8.5 | 25.68 | 17.92 | 27.46 |