Document Retrieval on MS MARCO MS300K (test)
57.6MRR@10GTR-Base
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GTR-BaseCategory=Dense(Brute-Force)2025.12 | 57.6 | 58.1 | 47.1 | 78.5 | 91.2 | — | |
| DDRO (TU)Docid strategy=TU (Title + URL)2025.04 | 50.07 | — | 38.24 | 74.01 | — | 66.46 | |
| DDRORetrieval paradigm=Generative retrieval, DocID design=Title + URL2025.04 | 50.07 | — | 38.24 | 74.01 | — | 66.46 | |
| SCI (Bert2Tower+ SymmAligner)Category=Dense(Brute-Force)2025.12 | 49.6 | 50.4 | 38.3 | 73.7 | 90.8 | — | |
| Bert2TowerCategory=Dense(Brute-Force)2025.12 | 48 | 48.8 | 37.2 | 71.4 | 89.3 | — | |
| LTRGRRetrieval paradigm=Generative retrieval, DocID design=Semantic ID2025.04 | 47.85 | — | 32.69 | 72.43 | — | 64.37 | |
| GenRRL (Sum)Docid strategy=Sum (document summary)2025.04 | 46.62 | — | 33.23 | 75.8 | — | 64.48 | |
| ROGER-UltronRetrieval paradigm=Generative retrieval, DocID design=Title + URL2025.04 | 46.35 | — | 33.07 | 75.13 | — | 63.93 | |
| GenRRL (TU)Docid strategy=TU (Title + URL)2025.04 | 45.93 | — | 33.01 | 74.91 | — | 63.62 | |
| DDRO (PQ)Docid strategy=PQ (Product Quantization)2025.04 | 45.76 | — | 32.92 | 73.02 | — | 64.36 | |
| DDRORetrieval paradigm=Generative retrieval, DocID design=Product Quantization2025.04 | 45.76 | — | 32.92 | 73.02 | — | 64.36 | |
| UltronRetrieval paradigm=Generative retrieval, DocID design=Product Quantization2025.04 | 45.35 | — | 31.55 | 73.14 | — | 63.98 | |
| ANCECategory=Dense(Brute-Force)2025.12 | 45.1 | 45.5 | 29.9 | 78.5 | 95.3 | — | |
| SCICategory=Dense(IVF-Flat), Alignment weight (λ)=0.32025.12 | 44.8 | 45.5 | 35 | 65.7 | 80.5 | — | |
| SPLADEv2Category=Sparse2025.12 | 44.3 | 45.2 | 32.8 | 77.9 | 95.6 | — | |
| ANCERetrieval paradigm=Dense retrieval2025.04 | 44.09 | — | 29.65 | 74.28 | — | 63.43 | |
| DPRRetrieval paradigm=Dense retrieval2025.04 | 43.41 | — | 29.08 | 73.13 | — | 62.75 | |
| UltronRetrieval paradigm=Generative retrieval, DocID design=Title + URL2025.04 | 42.53 | — | 29.82 | 68.31 | — | 60.39 | |
| MINDERRetrieval paradigm=Generative retrieval, DocID design=Semantic ID2025.04 | 42.51 | — | 29.98 | 71.92 | — | 58.37 | |
| UniCOILCategory=Sparse2025.12 | 42.5 | 43.5 | 28.4 | 76.6 | 95.1 | — | |
| DPRCategory=Dense(Brute-Force)2025.12 | 42.4 | 43.3 | 27.1 | 76.4 | 94.8 | — | |
| ROGER-NCIRetrieval paradigm=Generative retrieval, DocID design=Semantic ID2025.04 | 42.02 | — | 30.61 | 68.78 | — | 59.02 | |
| SCICategory=Dense(IVF-Flat), Alignment weight (λ)=0.42025.12 | 42 | 42.6 | 32.4 | 62 | 75.6 | — | |
| Bert2TowerCategory=Dense(IVF-Flat)2025.12 | 40.8 | 41.4 | 31.8 | 60.3 | 75 | — | |
| Sentence-T5Retrieval paradigm=Dense retrieval2025.04 | 40.69 | — | 27.27 | 72.15 | — | 58.91 | |
| NCIRetrieval paradigm=Generative retrieval, DocID design=Semantic ID2025.04 | 40.46 | — | 29.54 | 67.28 | — | 57.99 | |
| SCICategory=Dense(IVF-Flat), Alignment weight (λ)=0.22025.12 | 40 | 40.7 | 31.6 | 58 | 72.1 | — | |
| SCICategory=Dense(IVF-Flat), Alignment weight (λ)=0.52025.12 | 39.3 | 39.9 | 30.3 | 58.8 | 71.7 | — | |
| RepBERTRetrieval paradigm=Dense retrieval2025.04 | 38.48 | — | 25.25 | 69.18 | — | 58.41 | |
| DSI-QGRetrieval paradigm=Generative retrieval, DocID design=Semantic ID2025.04 | 38.45 | — | 28.82 | 62.26 | — | 50.74 | |
| SEALRetrieval paradigm=Generative retrieval, DocID design=N-grams2025.04 | 37.68 | — | 27.58 | 61.01 | — | 52.47 | |
| SCICategory=Dense(IVF-Flat), Alignment weight (λ)=0.12025.12 | 36.9 | 37.4 | 28.7 | 54.5 | 67.4 | — | |
| DocT5QueryRetrieval paradigm=Term-based retrieval2025.04 | 34.81 | — | 23.27 | 63.61 | — | 49.38 | |
| DSIRetrieval paradigm=Generative retrieval, DocID design=Semantic ID2025.04 | 33.92 | — | 25.74 | 53.84 | — | 43.58 | |
| BM25Retrieval paradigm=Term-based retrieval2025.04 | 29.24 | — | 18.94 | 55.07 | — | 42.82 | |
| BM25Category=Sparse2025.12 | 24.8 | 25.5 | 18.6 | 39.1 | 57.3 | — |