End-to-end Open-Domain Question Answering on NQ (test)
55.1Exact Match (EM)Qwen3-Emb-4B + Reranker
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Qwen3-Emb-4B + RerankerRetrieval method=Qwen3-Emb-4B + Reranker, Generator=T5Gemma2026.05 | 55.1 | — | — | — | — | |
| Qwen3-Emb-4BRetrieval method=Qwen3-Emb-4B, Generator=T5Gemma2026.05 | 54.5 | — | — | — | — | |
| UniK-QA#params=990M, Wikipedia-Tables=true2021.10 | 54 | — | — | — | — | |
| KG-FiD (large)#params=994M2021.10 | 53.4 | — | — | — | — | |
| BGERetrieval method=BGE, Generator=T5Gemma2026.05 | 52.2 | — | — | — | — | |
| FiD (large)#params=990M, implementation=ours2021.10 | 51.9 | — | — | — | — | |
| FiDReader Model Configuration=Large2021.01 | 51.4 | — | — | — | — | |
| Joint Top-kReader Model Configuration=Large2021.01 | 51.4 | — | — | — | — | |
| FiD (large)#params=990M, implementation=reported2021.10 | 51.4 | — | — | — | — | |
| INTRARetrieval method=INTRA, Generator=T5Gemma2026.05 | 51.2 | — | — | — | — | |
| Hybrid RAGRetrieval method=Hybrid RAG, Generator=T5Gemma2026.05 | 50.5 | — | — | — | — | |
| FastFiDNumber of retrieved passages=100, Number of selected sentences=2002024.08 | 50.17 | 148 | — | — | — | |
| HybridFiDNumber of retrieved passages=100, Number of selected sentences=2002024.08 | 50.14 | 513 | — | — | — | |
| FiDNumber of retrieved passages=100, Number of selected sentences=2002024.08 | 50.06 | 514 | — | — | — | |
| KG-FiD (base)#params=443M2021.10 | 49.6 | — | — | — | — | |
| Joint Top-kReader Model Configuration=Base2021.01 | 49.2 | — | — | — | — | |
| Joint Top K#params=440M2021.10 | 49.2 | — | — | — | — | |
| FiD (base)#params=440M, implementation=ours2021.10 | 48.8 | — | — | — | — | |
| MaxSimRetrieval method=MaxSim, Generator=T5Gemma2026.05 | 48.4 | — | — | — | — | |
| RIDER#params=626M2021.10 | 48.3 | — | — | — | — | |
| FiDReader Model Configuration=Base2021.01 | 48.2 | — | — | — | — | |
| FiD (base)#params=440M, implementation=reported2021.10 | 48.2 | — | — | — | — | |
| Individual Top-kReader Model Size=large2021.01 | 48.1 | — | — | — | — | |
| Individual Top-kReader Model Size=base2021.01 | 45.9 | — | — | — | — | |
| RECONSIDER#params=670M2021.10 | 45.5 | — | — | — | — | |
| RAGReader Model Size=large2021.01 | 44.5 | — | — | — | — | |
| RAG#params=626M2021.10 | 44.5 | — | — | — | — | |
| BM25Retrieval method=BM25, Generator=T5Gemma2026.05 | 43.4 | — | — | — | — | |
| DPRTraining=Single2020.04 | 41.5 | — | — | — | — | |
| DPRTraining=Multi2020.04 | 41.5 | — | — | — | — | |
| DPRReader Model Size=base2021.01 | 41.5 | — | — | — | — | |
| FiD-LightNumber of retrieved passages=100, Number of selected sentences=200, k=642024.08 | 40.91 | 201 | — | — | — | |
| REALMNewsTraining=Single, Pretraining=CC-News2020.04 | 40.4 | — | — | — | — | |
| REALMReader Model Size=base2021.01 | 40.4 | — | — | — | — | |
| REALM#params=330M2021.10 | 40.4 | — | — | — | — | |
| REALMWikiTraining=Single, Pretraining=Wikipedia2020.04 | 39.2 | — | — | — | — | |
| ECoRAGReader model=Gemma2, Compression strategy=RAG with 100 documents compressed2025.06 | 39.2 | — | — | 50.24 | 684 | |
| BM25+DPRTraining=Single2020.04 | 39 | — | — | — | — | |
| BM25+DPRTraining=Multi2020.04 | 38.8 | — | — | — | — | |
| CompActReader model=Gemma2, Compression strategy=RAG with 100 documents compressed2025.06 | 38.67 | — | — | 49.87 | 111 | |
| LongLLMLinguaReader model=Gemma2, Compression strategy=RAG with 100 documents compressed2025.06 | 37.67 | — | — | 49.4 | 1,411 | |
| RECOMP (extractive)Reader model=Gemma2, Compression strategy=RAG with 100 documents compressed2025.06 | 37.65 | — | — | 48.24 | 165 | |
| T5#params=11B2021.10 | 36.6 | — | — | — | — | |
| Qwen3-Emb-0.6BRetrieval method=Qwen3-Emb-0.6B, Generator=T5Gemma2026.05 | 36.4 | — | — | — | — | |
| TF-IDFRetrieval method=TF-IDF, Generator=T5Gemma2026.05 | 34.9 | — | — | — | — | |
| Graph-Retriever#params=110M2021.10 | 34.7 | — | — | — | — | |
| GraphRetrieverTraining=Single2020.04 | 34.5 | — | — | — | — | |
| ORQATraining=Single2020.04 | 33.3 | — | — | — | — | |
| ORQAReader Model Size=base2021.01 | 33.3 | — | — | — | — | |
| PathRetrieverTraining=Single2020.04 | 32.6 | — | — | — | — | |
| BM25Training=Single2020.04 | 32.6 | — | — | — | — | |
| Path-Retriever#params=445M2021.10 | 31.7 | — | — | — | — | |
| GPT-3#params=175B, config=few-shot2021.10 | 29.9 | — | — | — | — | |
| LLMLingua-2Reader model=Gemma2, Compression strategy=RAG with 100 documents compressed2025.06 | 28.56 | — | — | 38.95 | 1,403 | |
| HardEMTraining=Single2020.04 | 28.1 | — | — | — | — | |
| RECOMP (abstractive)Reader model=Gemma2, Compression strategy=RAG with 100 documents compressed2025.06 | 27.98 | — | — | 38 | 17 | |
| closed-bookReader model=Gemma2, Compression=None2025.06 | 27.84 | — | — | 38.35 | 0 | |
| LLMLinguaReader model=Gemma2, Compression strategy=RAG with 100 documents compressed2025.06 | 26.9 | — | — | 37.9 | 643 | |
| BM25+BERTTraining=Single2020.04 | 26.5 | — | — | — | — | |
| standard RAGReader model=Gemma2, Number of retrieved documents=1002025.06 | — | — | — | — | 14,260 |