{"paper":{"title":"Dense Passage Retrieval for Open-Domain Question Answering","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Dense vector embeddings from a dual-encoder model outperform BM25 by 9-19 percent in top-20 passage retrieval for open-domain question answering.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Barlas O\\u{g}uz, Danqi Chen, Ledell Wu, Patrick Lewis, Sergey Edunov, Sewon Min, Vladimir Karpukhin, Wen-tau Yih","submitted_at":"2020-04-10T04:53:17Z","abstract_excerpt":"Open-domain question answering relies on efficient passage retrieval to select candidate contexts, where traditional sparse vector space models, such as TF-IDF or BM25, are the de facto method. In this work, we show that retrieval can be practically implemented using dense representations alone, where embeddings are learned from a small number of questions and passages by a simple dual-encoder framework. When evaluated on a wide range of open-domain QA datasets, our dense retriever outperforms a strong Lucene-BM25 system largely by 9%-19% absolute in terms of top-20 passage retrieval accuracy,"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"our dense retriever outperforms a strong Lucene-BM25 system largely by 9%-19% absolute in terms of top-20 passage retrieval accuracy, and helps our end-to-end QA system establish new state-of-the-art on multiple open-domain QA benchmarks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That embeddings learned from a small number of questions and passages will generalize to the full distribution of queries and documents encountered at test time without additional adaptation.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Dense dual-encoder retrievers outperform BM25 by 9-19% absolute in top-20 passage retrieval accuracy across open-domain QA datasets and enable new state-of-the-art end-to-end QA results.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Dense vector embeddings from a dual-encoder model outperform BM25 by 9-19 percent in top-20 passage retrieval for open-domain question answering.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"34dcd6909a322924e26601d22cc89a2bb69a0ba33e8f2a386e0f66604458ec05"},"source":{"id":"2004.04906","kind":"arxiv","version":3},"verdict":{"id":"69d8be53-25bf-49ce-b023-2ebe61463f0e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T21:21:33.779649Z","strongest_claim":"our dense retriever outperforms a strong Lucene-BM25 system largely by 9%-19% absolute in terms of top-20 passage retrieval accuracy, and helps our end-to-end QA system establish new state-of-the-art on multiple open-domain QA benchmarks.","one_line_summary":"Dense dual-encoder retrievers outperform BM25 by 9-19% absolute in top-20 passage retrieval accuracy across open-domain QA datasets and enable new state-of-the-art end-to-end QA results.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That embeddings learned from a small number of questions and passages will generalize to the full distribution of queries and documents encountered at test time without additional adaptation.","pith_extraction_headline":"Dense vector embeddings from a dual-encoder model outperform BM25 by 9-19 percent in top-20 passage retrieval for open-domain question answering."},"references":{"count":113,"sample":[{"doi":"","year":null,"title":"Passage Re-ranking with","work_id":"fc2ddd97-10e7-40a7-933c-840da1cd2f01","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"2020 , booktitle =","work_id":"55e9beb5-adc8-4a04-8159-ea1fa498a605","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Relevance-guided Supervision for OpenQA with","work_id":"e109b3e7-d675-41ce-a9e1-1ef8465f9881","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"The probabilistic relevance framework:","work_id":"8fc0aab3-8cc7-4e70-9dc4-066311e5527d","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Learning to Retrieve Reasoning Paths over","work_id":"e513a414-13b7-4924-8112-988c627f3aea","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":113,"snapshot_sha256":"bcf46dcd452c8eedc2f350ab9bd78dfcfcbaa00730946212898e387189d96408","internal_anchors":8},"formal_canon":{"evidence_count":1,"snapshot_sha256":"916c72306ca4ee458b29fbf2bcdfea177e7ae5eff4f5bf83c2cdeee0ce2dae6b"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}