{"paper":{"title":"Teaching language models to support answers with verified quotes","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A 280 billion parameter model can be trained to answer questions with specific cited evidence from documents and to abstain when uncertain.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Francis Song, Geoffrey Irving, Jacob Menick, John Aslanides, Lucy Campbell-Gillingham, Maja Trebacz, Martin Chadwick, Mia Glaese, Nat McAleese, Susannah Young, Vladimir Mikulik","submitted_at":"2022-03-21T17:26:29Z","abstract_excerpt":"Recent large language models often answer factual questions correctly. But users can't trust any given claim a model makes without fact-checking, because language models can hallucinate convincing nonsense. In this work we use reinforcement learning from human preferences (RLHP) to train \"open-book\" QA models that generate answers whilst also citing specific evidence for their claims, which aids in the appraisal of correctness. Supporting evidence is drawn from multiple documents found via a search engine, or from a single user-provided document. Our 280 billion parameter model, GopherCite, is"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our 280 billion parameter model, GopherCite, is able to produce answers with high quality supporting evidence and abstain from answering when unsure. The model's response is found to be high-quality 80% of the time on this Natural Questions subset, and 67% of the time on the ELI5 subset. Abstaining from the third of questions for which it is most unsure improves performance to 90% and 80% respectively.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That human raters' preferences for 'high quality supporting evidence' during RLHP training generalize to produce reliable citations and that the model's internal uncertainty signal for abstention is well-calibrated without introducing new biases.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GopherCite produces answers with supporting evidence citations, rated high-quality 80% of the time on Natural Questions and 67% on ELI5, improving to 90% and 80% with abstention on uncertain questions.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A 280 billion parameter model can be trained to answer questions with specific cited evidence from documents and to abstain when uncertain.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0bfbea845d9b4ea5c26a8c4ad7ed798d6d75e5d37e0a6c3c78e702e9d62020fd"},"source":{"id":"2203.11147","kind":"arxiv","version":1},"verdict":{"id":"9612394c-192c-4c1f-a9f9-d2f981f7f4f1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T11:41:27.964902Z","strongest_claim":"Our 280 billion parameter model, GopherCite, is able to produce answers with high quality supporting evidence and abstain from answering when unsure. The model's response is found to be high-quality 80% of the time on this Natural Questions subset, and 67% of the time on the ELI5 subset. Abstaining from the third of questions for which it is most unsure improves performance to 90% and 80% respectively.","one_line_summary":"GopherCite produces answers with supporting evidence citations, rated high-quality 80% of the time on Natural Questions and 67% on ELI5, improving to 90% and 80% with abstention on uncertain questions.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That human raters' preferences for 'high quality supporting evidence' during RLHP training generalize to produce reliable citations and that the model's internal uncertainty signal for abstention is well-calibrated without introducing new biases.","pith_extraction_headline":"A 280 billion parameter model can be trained to answer questions with specific cited evidence from documents and to abstain when uncertain."},"references":{"count":14,"sample":[{"doi":"10.1145/3041021.3053375","year":2018,"title":"ISBN 9781450349147","work_id":"f9659ba7-d3b9-47d1-a957-64fed829351f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1177/0894439317715434","year":2018,"title":"road draft tube","work_id":"be985a9d-c962-4bf8-91fa-284ca394a293","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"{url} • {claim} See this fragment from \"{title}\"[1]: {quote}","work_id":"5aad03f9-dc2e-4995-b1ce-97c186899a82","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"{quote}\" Source:","work_id":"23284550-9063-4b4d-90d9-825a613e8ccd","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"What happens if you smash a mirror?","work_id":"712292c7-2504-4989-b456-03e962093e6c","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":14,"snapshot_sha256":"3aa19ac14cbd371c1e693ce659656e65f10ba455726feb7f80c2706295bbbc85","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}