Multimodal Question Answering on WebQA Image-grounded (test, n=2,511)
69.4F1 ScoreLatent Memory
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Latent Memoryk=5, Generator VLM=LLaVA-1.5-13B, #Tok=822026.06 | 69.4 | 44.2 | 91.2 | |
| Latent Memoryk=2, Generator VLM=LLaVA-1.5-13B, #Tok=522026.06 | 56.5 | 39.5 | 74.7 | |
| Nemo-Emb-1Bk=5, Generator VLM=LLaVA-1.5-13B, #Tok=18852026.06 | 53 | 37.2 | 81.4 | |
| Dense Retrievalk=5, Generator VLM=LLaVA-1.5-13B, #Tok=16452026.06 | 49.8 | 36.6 | 67.1 | |
| Qwen3-VL-Emb-8Bk=5, Generator VLM=LLaVA-1.5-13B, #Tok=9572026.06 | 34.1 | 30.3 | 49.7 | |
| Nemo-Emb-1Bk=2, Generator VLM=LLaVA-1.5-13B, #Tok=8922026.06 | 32.9 | 30.6 | 64.8 | |
| BM25 Retrievalk=5, Generator VLM=LLaVA-1.5-13B, #Tok=9322026.06 | 32.6 | 29.5 | 37.9 | |
| Latent Memoryk=1, Generator VLM=LLaVA-1.5-13B, #Tok=422026.06 | 32 | 28.7 | 56.6 | |
| Dense Retrievalk=2, Generator VLM=LLaVA-1.5-13B, #Tok=8142026.06 | 30 | 30.1 | 53.2 | |
| Qwen3-VL-Emb-8Bk=2, Generator VLM=LLaVA-1.5-13B, #Tok=4652026.06 | 20.1 | 24.5 | 32.9 | |
| Nemo-Emb-1Bk=1, Generator VLM=LLaVA-1.5-13B, #Tok=5072026.06 | 19.9 | 25.2 | 48.7 | |
| BM25 Retrievalk=2, Generator VLM=LLaVA-1.5-13B, #Tok=4762026.06 | 19.2 | 23.4 | 28.3 | |
| Dense Retrievalk=1, Generator VLM=LLaVA-1.5-13B, #Tok=4762026.06 | 18.9 | 24.5 | 39.7 | |
| Qwen3-VL-Emb-8Bk=1, Generator VLM=LLaVA-1.5-13B, #Tok=2842026.06 | 15.1 | 22 | 24.1 | |
| BM25 Retrievalk=1, Generator VLM=LLaVA-1.5-13B, #Tok=2952026.06 | 14.5 | 21.5 | 21.8 | |
| Full ContextGenerator VLM=LLaVA-1.5-13B, #Tok=116552026.06 | 0 | 0 | — |