Multimodal Retrieval-Augmented Generation on CRAG-MM Egocentric
-11.7TruthfulnessPixSearch-13B
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| PixSearch-13BBackbone=LLaVA-13B, Search Strategy=Full2026.01 | -11.7 | 33 | 22.3 | 44.7 | |
| Llama-3.2-11BBackbone=Llama-3.2-11B-Vision, Search Strategy=Whole Image2026.01 | -17 | 31 | 21 | 48 | |
| Llama-3.2-11BBackbone=Llama-3.2-11B-Vision, Search Strategy=No Search2026.01 | -22.5 | 21 | 35.5 | 43.5 | |
| PixSearch-13BBackbone=LLaVA-13B, Search Strategy=Whole Image2026.01 | -27.6 | 26.7 | 19 | 54.3 | |
| GroundedSAM+LLaVA-13BBackbone=LLaVA-13B, Search Strategy=Pipeline Method (Region-Level)2026.01 | -41.7 | 18.1 | 22 | 59.8 | |
| PLUM-13BBackbone=LLaVA-13B2026.01 | -43.2 | 20.5 | 15.8 | 63.7 | |
| PixSearch-13BBackbone=LLaVA-13B, Search Strategy=No Search2026.01 | -47.8 | 15.7 | 20.8 | 63.5 | |
| PixSearch-13BBackbone=LLaVA-13B, Search Strategy=Text Query2026.01 | -52.5 | 12.2 | 20.1 | 67.7 | |
| LLaVA-13BBackbone=LLaVA-13B, Search Strategy=None2026.01 | -72.8 | 8.1 | 11 | 80.9 |