Video Question Answering on EventBench 60-1800s (test)
62.28Contextual AccuracyInternVL3.5-8B + EC-RAG
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| InternVL3.5-8B + EC-RAGBackbone=InternVL3.5-8B, Augmentation=EC-RAG2026.05 | 62.28 | 46.67 | 57.27 | 55.97 | |
| Qwen3-VL-8B + EC-RAGBackbone=Qwen3-VL-8B, Augmentation=EC-RAG2026.05 | 61.77 | 48.33 | 53.3 | 55.31 | |
| VideoLlama3-7B + EC-RAGBackbone=VideoLlama3-7B, Augmentation=EC-RAG2026.05 | 61.01 | 49.67 | 55.51 | 55.97 | |
| InternVL3.5-8BBackbone=InternVL3.5-8B2026.05 | 56.96 | 45 | 60.35 | 53.9 | |
| VideoLlama3-7BBackbone=VideoLlama3-7B2026.05 | 56.71 | 42 | 55.95 | 51.74 | |
| Qwen3-VL-8BBackbone=Qwen3-VL-8B2026.05 | 53.16 | 45.67 | 59.91 | 52.39 | |
| GPT-4o (Closed-source)Access=Closed-source2026.05 | 50.13 | 37.33 | 63.44 | 53.33 | |
| Gemini1.5-pro (Closed-source)Access=Closed-source2026.05 | 32.15 | 38.67 | 52.86 | 43.24 |