Visual Document Retrieval on VisDoc OOD (NDCG@5)
73.5NDCG@5NemColEmbed V2 8B
Evaluation results
| Method | Links | |
|---|---|---|
| NemColEmbed V2 8BSize=8B, Late Interaction=true, Neural Query Encoder=true2026.05 | 73.5 | |
| GME-Qwen2-7BSize=7B, Late Interaction=true2026.05 | 69.3 | |
| Qwen3-VL-Emb-2BSize=2B, Late Interaction=true2026.05 | 69.2 | |
| ColPaliSize=3B, Late Interaction=true, Neural Query Encoder=true2026.05 | 66.5 | |
| ColModernVBERTSize=0.25B, Late Interaction=true, Neural Query Encoder=true2026.05 | 66.1 | |
| V-SPLADE + BiModernVBERT + BM25 (unstr.)Size=0.25/0.25B, Neural Query Encoder=true2026.05 | 64 | |
| V-SPLADE + BiModernVBERTSize=0.25/0.25B, Neural Query Encoder=true2026.05 | 63.8 | |
| V-SPLADE + BM25 (unstr.)Size=0.25B2026.05 | 62.2 | |
| V-SPLADE qualitySize=0.25B, FLOPs=1.92026.05 | 61.7 | |
| E5-VSize=8B, Late Interaction=true2026.05 | 60.8 | |
| BGE-M3 (caption)Size=0.57B, Late Interaction=true2026.05 | 60.7 | |
| V-SPLADE efficientSize=0.25B, FLOPs=1.12026.05 | 59.5 | |
| BM25 (caption)FLOPs=1.32026.05 | 58 | |
| ColFlorSize=0.17B, Late Interaction=true, Neural Query Encoder=true2026.05 | 57.3 | |
| VLM2VecSize=4B, Late Interaction=true2026.05 | 54 | |
| BiModernVBERTSize=0.25B, Neural Query Encoder=true2026.05 | 53.4 | |
| BGE-M3 (unstr.)Size=0.57B, Late Interaction=true2026.05 | 52.1 | |
| BM25 (unstr.)FLOPs=1.42026.05 | 51.2 | |
| Jina CLIP v2Size=0.9B, Neural Query Encoder=true2026.05 | 47.2 | |
| SigLIP2-LSize=0.9B, Neural Query Encoder=true2026.05 | 34.8 | |
| JSDO-SparseSize=0.25B, Neural Query Encoder=true, FLOPs=18142026.05 | 29.7 | |
| D2SSize=0.25B, Neural Query Encoder=true, FLOPs=14.72026.05 | 9.6 |