Information Retrieval on MS MARCO (test)
0.714NDCG@10Ours
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| OursBackbone=Qwen-8B, Training Protocol=supervised with relevance labels2026.02 | 0.714 | 20 | — | |
| Ours@30%Backbone=Qwen-8B, Training Protocol=supervised with relevance labels, Fixed Dimension Ratio=30%2026.02 | 0.709 | 30 | — | |
| DIMEBackbone=OpenAI, Training Protocol=PRF-based2026.02 | 0.702 | 10 | — | |
| EclipseBackbone=Qwen-4B, Training Protocol=PRF-based2026.02 | 0.702 | 28 | — | |
| DIMEBackbone=Qwen-4B, Training Protocol=PRF-based2026.02 | 0.701 | 14 | — | |
| EclipseBackbone=OpenAI, Training Protocol=PRF-based2026.02 | 0.7 | 32 | — | |
| OursBackbone=OpenAI, Training Protocol=supervised with relevance labels2026.02 | 0.7 | 34 | — | |
| OursBackbone=Qwen-4B, Training Protocol=supervised with relevance labels2026.02 | 0.699 | 20 | — | |
| AdapterBackbone=Qwen-8B, Training Protocol=supervised with relevance labels2026.02 | 0.698 | 100 | — | |
| Ours@30%Backbone=OpenAI, Training Protocol=supervised with relevance labels, Fixed Dimension Ratio=30%2026.02 | 0.696 | 30 | — | |
| Ours@30%Backbone=Qwen-4B, Training Protocol=supervised with relevance labels, Fixed Dimension Ratio=30%2026.02 | 0.693 | 30 | — | |
| CutoffBackbone=OpenAI, Training Protocol=fully unsupervised2026.02 | 0.69 | 74 | — | |
| NormBackbone=OpenAI, Training Protocol=fully unsupervised2026.02 | 0.688 | 10 | — | |
| NormBackbone=Qwen-4B, Training Protocol=fully unsupervised2026.02 | 0.687 | 42 | — | |
| BaselineBackbone=Qwen-4B, Training Protocol=fully unsupervised2026.02 | 0.683 | 100 | — | |
| CutoffBackbone=Qwen-4B, Training Protocol=fully unsupervised2026.02 | 0.683 | 100 | — | |
| AdapterBackbone=Qwen-4B, Training Protocol=supervised with relevance labels2026.02 | 0.682 | 100 | — | |
| BaselineBackbone=OpenAI, Training Protocol=fully unsupervised2026.02 | 0.681 | 100 | — | |
| AdapterBackbone=OpenAI, Training Protocol=supervised with relevance labels2026.02 | 0.68 | 100 | — | |
| EclipseBackbone=Qwen-8B, Training Protocol=PRF-based2026.02 | 0.668 | 84 | — | |
| DIMEBackbone=Qwen-8B, Training Protocol=PRF-based2026.02 | 0.664 | 40 | — | |
| NormBackbone=Qwen-8B, Training Protocol=fully unsupervised2026.02 | 0.653 | 62 | — | |
| CutoffBackbone=Qwen-8B, Training Protocol=fully unsupervised2026.02 | 0.651 | 52 | — | |
| BaselineBackbone=Qwen-8B, Training Protocol=fully unsupervised2026.02 | 0.646 | 100 | — | |
| EclipseBackbone=L2V-Mistral, Training Protocol=PRF-based2026.02 | 0.641 | 86 | — | |
| DIMEBackbone=L2V-Mistral, Training Protocol=PRF-based2026.02 | 0.638 | 70 | — | |
| OursBackbone=L2V-Mistral, Training Protocol=supervised with relevance labels2026.02 | 0.638 | 24 | — | |
| Ours@30%Backbone=L2V-Mistral, Training Protocol=supervised with relevance labels, Fixed Dimension Ratio=30%2026.02 | 0.634 | 30 | — | |
| OursBackbone=GritLM, Training Protocol=supervised with relevance labels2026.02 | 0.632 | 40 | — | |
| Ours@30%Backbone=GritLM, Training Protocol=supervised with relevance labels, Fixed Dimension Ratio=30%2026.02 | 0.631 | 30 | — | |
| CutoffBackbone=L2V-Mistral, Training Protocol=fully unsupervised2026.02 | 0.623 | 76 | — | |
| AdapterBackbone=L2V-Mistral, Training Protocol=supervised with relevance labels2026.02 | 0.621 | 100 | — | |
| NormBackbone=L2V-Mistral, Training Protocol=fully unsupervised2026.02 | 0.617 | 40 | — | |
| BaselineBackbone=L2V-Mistral, Training Protocol=fully unsupervised2026.02 | 0.613 | 100 | — | |
| OursBackbone=Qwen-0.6B, Training Protocol=supervised with relevance labels2026.02 | 0.609 | 44 | — | |
| AdapterBackbone=Qwen-0.6B, Training Protocol=supervised with relevance labels2026.02 | 0.607 | 100 | — | |
| EclipseBackbone=Qwen-0.6B, Training Protocol=PRF-based2026.02 | 0.602 | 18 | — | |
| Ours@30%Backbone=Qwen-0.6B, Training Protocol=supervised with relevance labels, Fixed Dimension Ratio=30%2026.02 | 0.601 | 30 | — | |
| OursBackbone=L2V-LLaMA, Training Protocol=supervised with relevance labels2026.02 | 0.6 | 26 | — | |
| DIMEBackbone=Qwen-0.6B, Training Protocol=PRF-based2026.02 | 0.595 | 84 | — | |
| AdapterBackbone=GritLM, Training Protocol=supervised with relevance labels2026.02 | 0.593 | 100 | — | |
| Ours@30%Backbone=L2V-LLaMA, Training Protocol=supervised with relevance labels, Fixed Dimension Ratio=30%2026.02 | 0.593 | 30 | — | |
| EclipseBackbone=L2V-LLaMA, Training Protocol=PRF-based2026.02 | 0.588 | 50 | — | |
| DIMEBackbone=L2V-LLaMA, Training Protocol=PRF-based2026.02 | 0.582 | 64 | — | |
| AdapterBackbone=L2V-LLaMA, Training Protocol=supervised with relevance labels2026.02 | 0.576 | 100 | — | |
| NormBackbone=Qwen-0.6B, Training Protocol=fully unsupervised2026.02 | 0.575 | 52 | — | |
| NormBackbone=L2V-LLaMA, Training Protocol=fully unsupervised2026.02 | 0.569 | 48 | — | |
| DIMEBackbone=GritLM, Training Protocol=PRF-based2026.02 | 0.566 | 64 | — | |
| CutoffBackbone=L2V-LLaMA, Training Protocol=fully unsupervised2026.02 | 0.565 | 94 | — | |
| EclipseBackbone=GritLM, Training Protocol=PRF-based2026.02 | 0.565 | 76 | — | |
| BaselineBackbone=Qwen-0.6B, Training Protocol=fully unsupervised2026.02 | 0.562 | 100 | — | |
| BaselineBackbone=L2V-LLaMA, Training Protocol=fully unsupervised2026.02 | 0.562 | 100 | — | |
| CutoffBackbone=Qwen-0.6B, Training Protocol=fully unsupervised2026.02 | 0.562 | 100 | — | |
| CutoffBackbone=GritLM, Training Protocol=fully unsupervised2026.02 | 0.554 | 54 | — | |
| NormBackbone=GritLM, Training Protocol=fully unsupervised2026.02 | 0.537 | 86 | — | |
| BaselineBackbone=GritLM, Training Protocol=fully unsupervised2026.02 | 0.536 | 100 | — | |
| AligneRRetrieval Type=Late-interaction dense retrieval2026.05 | — | — | 51.8 | |
| CITADELRetrieval Type=Late-interaction dense retrieval2026.05 | — | — | 15.7 | |
| COILRetrieval Type=Late-interaction dense retrieval2026.05 | — | — | 12.6 | |
| ColBERTRetrieval Type=Late-interaction dense retrieval2026.05 | — | — | 57.3 | |
| ColBERTv2Retrieval Type=Late-interaction dense retrieval2026.05 | — | — | 56.9 | |
| PLAIDRetrieval Type=Late-interaction dense retrieval2026.05 | — | — | 37.1 | |
| Splade-v2Retrieval Type=Learned sparse retrieval2026.05 | — | — | 16.3 | |
| Splade-v3Retrieval Type=Learned sparse retrieval2026.05 | — | — | 16.6 | |
| SSR-CLSRetrieval Type=Late-interaction sparse retrieval, acceleration_strategy=SSR++2026.05 | — | — | 19.5 | |
| SSR-tokRetrieval Type=Late-interaction sparse retrieval, acceleration_strategy=SSR++2026.05 | — | — | 17.5 | |
| XTRRetrieval Type=Late-interaction dense retrieval2026.05 | — | — | 33.4 |