{"paper":{"title":"REALM: Retrieval-Augmented Language Model Pre-Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Language models pre-trained with an integrated retriever over a document corpus outperform prior methods on open-domain question answering by 4 to 16 percent.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Kelvin Guu, Kenton Lee, Ming-Wei Chang, Panupong Pasupat, Zora Tung","submitted_at":"2020-02-10T18:40:59Z","abstract_excerpt":"Language model pre-training has been shown to capture a surprising amount of world knowledge, crucial for NLP tasks such as question answering. However, this knowledge is stored implicitly in the parameters of a neural network, requiring ever-larger networks to cover more facts.\n  To capture knowledge in a more modular and interpretable way, we augment language model pre-training with a latent knowledge retriever, which allows the model to retrieve and attend over documents from a large corpus such as Wikipedia, used during pre-training, fine-tuning and inference. For the first time, we show h"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We demonstrate the effectiveness of Retrieval-Augmented Language Model pre-training (REALM) by fine-tuning on the challenging task of Open-domain Question Answering (Open-QA). We compare against state-of-the-art models for both explicit and implicit knowledge storage on three popular Open-QA benchmarks, and find that we outperform all previous methods by a significant margin (4-16% absolute accuracy).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That back-propagation through a retrieval step over millions of documents is numerically stable and provides a useful unsupervised learning signal for the retriever parameters.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"REALM augments language-model pre-training with an unsupervised retriever over Wikipedia documents and reports 4-16% absolute gains on open-domain QA benchmarks over prior implicit and explicit knowledge methods.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Language models pre-trained with an integrated retriever over a document corpus outperform prior methods on open-domain question answering by 4 to 16 percent.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"fd05d3e400545758fdf5d7a79da9bff47de60a2fdeaaa7ce976fb7785f375b58"},"source":{"id":"2002.08909","kind":"arxiv","version":1},"verdict":{"id":"0496c68e-e097-4692-837a-27d9f626107a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T09:52:42.388463Z","strongest_claim":"We demonstrate the effectiveness of Retrieval-Augmented Language Model pre-training (REALM) by fine-tuning on the challenging task of Open-domain Question Answering (Open-QA). We compare against state-of-the-art models for both explicit and implicit knowledge storage on three popular Open-QA benchmarks, and find that we outperform all previous methods by a significant margin (4-16% absolute accuracy).","one_line_summary":"REALM augments language-model pre-training with an unsupervised retriever over Wikipedia documents and reports 4-16% absolute gains on open-domain QA benchmarks over prior implicit and explicit knowledge methods.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That back-propagation through a retrieval step over millions of documents is numerically stable and provides a useful unsupervised learning signal for the retriever parameters.","pith_extraction_headline":"Language models pre-trained with an integrated retriever over a document corpus outperform prior methods on open-domain question answering by 4 to 16 percent."},"references":{"count":20,"sample":[{"doi":"","year":1911,"title":"arXiv preprint arXiv:1911.10470 , year=","work_id":"bba0e5b4-dccd-4c07-86fb-2dae46efdce9","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Neural Machine Translation by Jointly Learning to Align and Translate","work_id":"d831e763-d530-4029-a65c-ac595d82cb2a","ref_index":2,"cited_arxiv_id":"1409.0473","is_internal_anchor":true},{"doi":"","year":2013,"title":"Semantic parsing on freebase from question-answer pairs","work_id":"55a52a96-0321-4fc0-a8f5-eccec466bc2d","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding","work_id":"ed240a10-5b19-406c-baa5-30803f465785","ref_index":4,"cited_arxiv_id":"1810.04805","is_internal_anchor":true},{"doi":"","year":null,"title":"Neural Turing Machines","work_id":"0a5ce53c-9670-42b9-8be5-386de7eed50c","ref_index":5,"cited_arxiv_id":"1410.5401","is_internal_anchor":true}],"resolved_work":20,"snapshot_sha256":"217ebf269247a333e779ed5dbbf87e45e2bf0a2c4d5f6cbb1950fc955a6166ce","internal_anchors":11},"formal_canon":{"evidence_count":3,"snapshot_sha256":"921de4a55a8393b7609ce9eec082d809fbe45062a568b445ccd3f0a7c006572d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}