Multimodal Question Answering on WebQA Image
13.8F1 ScoreFull Context
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Full ContextGenerator=Gemma-3-12B-Instruct, Retrieval Method=Full-context reference2026.06 | 13.8 | 18 | — | 5,856 | |
| Latent MemoryGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=12026.06 | 11.7 | 18.7 | 52.5 | 38 | |
| Nemo-EmbGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=12026.06 | 11.3 | 17.8 | 48.8 | 255 | |
| Latent MemoryGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=22026.06 | 11.3 | 18 | 70.1 | 47 | |
| Latent MemoryGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=52026.06 | 11.3 | 17.5 | 89 | 74 | |
| Nemo-EmbGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=52026.06 | 11.2 | 17.3 | 81.5 | 948 | |
| Nemo-EmbGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=22026.06 | 11.1 | 17 | 64.7 | 444 | |
| Dense RetrievalGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=12026.06 | 10.6 | 15.8 | 39.7 | 241 | |
| Dense RetrievalGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=52026.06 | 10.6 | 14.7 | 67.1 | 848 | |
| Dense RetrievalGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=22026.06 | 10.3 | 14.6 | 53.1 | 411 | |
| Qwen3-VL-8BGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=52026.06 | 9.5 | 12.4 | 49.6 | 570 | |
| Qwen3-VL-8BGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=22026.06 | 9.1 | 11.1 | 32.6 | 270 | |
| BM25 RetrievalGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=52026.06 | 8.8 | 10.5 | 37.9 | 559 | |
| Qwen3-VL-8BGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=12026.06 | 8.6 | 11.6 | 24 | 164 | |
| BM25 RetrievalGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=22026.06 | 8.3 | 9.5 | 28.3 | 274 | |
| BM25 RetrievalGenerator=Gemma-3-12B-Instruct, Retrieval count (k)=12026.06 | 8.2 | 9.2 | 21.8 | 167 |