Multimodal Fact-Level Attribution on Video-MMMU 1.0 (sampled examples)
86.8AccuracyGemini-3-Flash
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Gemini-3-FlashVariant=Base2026.02 | 86.8 | — | — | — | |
| Gemini-3-FlashVariant=+ Post-Hoc Attribution2026.02 | 86.8 | 87.9 | 47.2 | 44.1 | |
| Gemini-3-FlashVariant=+ Citation2026.02 | 86 | 88.2 | 64.5 | 56.9 | |
| Gemini-3-ProVariant=+ Citation2026.02 | 86 | 63.4 | 67.3 | 41.8 | |
| Gemini-3-ProVariant=Base2026.02 | 85.3 | — | — | — | |
| Gemini-3-ProVariant=+ Post-Hoc Attribution2026.02 | 85.3 | 68 | 43.7 | 36.9 | |
| Gemini-2.5-FlashVariant=+ Citation2026.02 | 84.9 | 63 | 63.4 | 41.5 | |
| Gemini-2.5-FlashVariant=Base2026.02 | 84.2 | — | — | — | |
| Gemini-2.5-FlashVariant=+ Post-Hoc Attribution2026.02 | 84.2 | 73.8 | 44.9 | 38 | |
| Qwen-3-VL-ThinkingCategory=Vision-Language Only, Variant=+ Citation2026.02 | 60 | 23.2 | 15.1 | 7.6 | |
| Qwen-3-VL-InstructCategory=Vision-Language Only, Variant=+ Citation2026.02 | 55 | 30.2 | 40.1 | 17.5 | |
| Qwen-Omni-ThinkingVariant=Base2026.02 | 53 | — | — | — | |
| Qwen-Omni-ThinkingVariant=+ Post-Hoc Attribution2026.02 | 53 | 76.3 | 16.8 | 12.8 | |
| Qwen-3-VL-InstructCategory=Vision-Language Only, Variant=Base2026.02 | 53 | — | — | — | |
| Qwen-3-VL-InstructCategory=Vision-Language Only, Variant=+ Post-Hoc Attribution2026.02 | 53 | 93.4 | 44.6 | 42.3 | |
| Qwen-Omni-ThinkingVariant=+ Citation2026.02 | 51 | 36.3 | 7.6 | 4.8 | |
| Qwen-3-VL-ThinkingCategory=Vision-Language Only, Variant=Base2026.02 | 51 | — | — | — | |
| Qwen-3-VL-ThinkingCategory=Vision-Language Only, Variant=+ Post-Hoc Attribution2026.02 | 51 | 54.3 | 31.5 | 18.9 | |
| Molmo2Category=Vision-Language Only, Variant=Base2026.02 | 50.5 | — | — | — | |
| Molmo2Category=Vision-Language Only, Variant=+ Post-Hoc Attribution2026.02 | 50.5 | 66.4 | 15 | 11.4 | |
| Qwen-Omni-InstructVariant=Base2026.02 | 45 | — | — | — | |
| Qwen-Omni-InstructVariant=+ Post-Hoc Attribution2026.02 | 45 | 95.1 | 17.9 | 17.6 | |
| Molmo2Category=Vision-Language Only, Variant=+ Citation2026.02 | 44.3 | 82.6 | 21.4 | 19.3 | |
| Qwen-Omni-InstructVariant=+ Citation2026.02 | 40 | 34.6 | 21.8 | 9.8 | |
| Gemini-2.5-Flash2026.02 | — | 57.1 | 32.5 | 21.8 | |
| Gemini-3-Pro2026.02 | — | 59.7 | 24.8 | 16.3 | |
| Qwen3-Omni-Instruct2026.02 | — | 35.9 | 14.9 | 5.6 | |
| Qwen3-Omni-Thinking2026.02 | — | 45 | 23.4 | 21.8 |