{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2021:UKSLWPQT2Z3UNZ3JUCL37JYPQF","short_pith_number":"pith:UKSLWPQT","canonical_record":{"source":{"id":"2109.10862","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2021-09-22T17:34:18Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"612ddce1253fc4e6a3ead618e152978b1fbe6ff31ad268ab30b5b82b9e65fc6e","abstract_canon_sha256":"ecaab0f0204b68cab469b5a1be1cb0e5693ea4b755132f7c575933fc96022bdb"},"schema_version":"1.0"},"canonical_sha256":"a2a4bb3e13d67746e769a097bfa70f8148c475fb3b8ffa86c331751c38e0255a","source":{"kind":"arxiv","id":"2109.10862","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2109.10862","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"arxiv_version","alias_value":"2109.10862v2","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2109.10862","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"pith_short_12","alias_value":"UKSLWPQT2Z3U","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"UKSLWPQT2Z3UNZ3J","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"UKSLWPQT","created_at":"2026-05-18T12:33:33Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2021:UKSLWPQT2Z3UNZ3JUCL37JYPQF","target":"record","payload":{"canonical_record":{"source":{"id":"2109.10862","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2021-09-22T17:34:18Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"612ddce1253fc4e6a3ead618e152978b1fbe6ff31ad268ab30b5b82b9e65fc6e","abstract_canon_sha256":"ecaab0f0204b68cab469b5a1be1cb0e5693ea4b755132f7c575933fc96022bdb"},"schema_version":"1.0"},"canonical_sha256":"a2a4bb3e13d67746e769a097bfa70f8148c475fb3b8ffa86c331751c38e0255a","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:13.697225Z","signature_b64":"5n+v4tOAmms5ajGNA+c42Iub3gqC6Ixo9pa1IxIQ5GTkJJ2wJqbEUKJxZxHzfgfWjevCCzfZ81R5Zg015HEwDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a2a4bb3e13d67746e769a097bfa70f8148c475fb3b8ffa86c331751c38e0255a","last_reissued_at":"2026-05-17T23:38:13.696569Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:13.696569Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2109.10862","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:13Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"4jwGJPKKTqXbGYkbSXlea4VWiCR9SfAdnptPQf1FQv4M69KwoeXHXnvxLdzGjZ8X6gpwbnXukL03/yQEe+UEAg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-23T19:10:45.574392Z"},"content_sha256":"de5808fbf6703c24c07c86c3bfff303c374f0c66c8d616868b39aad5d7b0f36a","schema_version":"1.0","event_id":"sha256:de5808fbf6703c24c07c86c3bfff303c374f0c66c8d616868b39aad5d7b0f36a"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2021:UKSLWPQT2Z3UNZ3JUCL37JYPQF","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Recursively Summarizing Books with Human Feedback","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Recursive decomposition lets models summarize entire books after humans give feedback only on short sections.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Daniel M. Ziegler, Jan Leike, Jeff Wu, Long Ouyang, Nisan Stiennon, Paul Christiano, Ryan Lowe","submitted_at":"2021-09-22T17:34:18Z","abstract_excerpt":"A major challenge for scaling machine learning is training models to perform tasks that are very difficult or time-consuming for humans to evaluate. We present progress on this problem on the task of abstractive summarization of entire fiction novels. Our method combines learning from human feedback with recursive task decomposition: we use models trained on smaller parts of the task to assist humans in giving feedback on the broader task. We collect a large volume of demonstrations and comparisons from human labelers, and fine-tune GPT-3 using behavioral cloning and reward modeling to do summ"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our resulting model generates sensible summaries of entire books, even matching the quality of human-written summaries in a few cases (~5% of books). We achieve state-of-the-art results on the recent BookSum dataset for book-length summarization.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That summaries of summaries retain enough information and fidelity for the final output to remain faithful to the original book when humans never see the full text.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Recursive decomposition plus human feedback lets GPT-3 produce book-length summaries that reach human quality on a few cases and set new records on BookSum and NarrativeQA.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Recursive decomposition lets models summarize entire books after humans give feedback only on short sections.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ac2d3692783a95b8708e95936c8d8684c6c8e48ee69de27882d9644785e03bea"},"source":{"id":"2109.10862","kind":"arxiv","version":2},"verdict":{"id":"4593b665-e34c-45c1-8e5b-878911626631","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T15:22:19.704135Z","strongest_claim":"Our resulting model generates sensible summaries of entire books, even matching the quality of human-written summaries in a few cases (~5% of books). We achieve state-of-the-art results on the recent BookSum dataset for book-length summarization.","one_line_summary":"Recursive decomposition plus human feedback lets GPT-3 produce book-length summaries that reach human quality on a few cases and set new records on BookSum and NarrativeQA.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That summaries of summaries retain enough information and fidelity for the final output to remain faithful to the original book when humans never see the full text.","pith_extraction_headline":"Recursive decomposition lets models summarize entire books after humans give feedback only on short sections."},"references":{"count":12,"sample":[{"doi":"","year":null,"title":"This subtask can be decomposed even further if necessary","work_id":"0edfbdfc-d42f-42ee-85ca-86cabf1cf023","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"answer_directly, which returns an actual answer to the task, synthesizing the answers to subtasks In general, both decompose_if_needed and answer_directly could be learned and implemented by an ML mod","work_id":"a86a2eb4-2be7-457f-b24e-1f5157bb1f02","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"So gratuitously including small details is generally penalized, and omitting important details is also penalized","work_id":"18f5ae1f-6b0b-4c05-a46f-7f852267c26b","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Accuracy: All information in the summary should faithfully reflect the original passage","work_id":"0c69f10e-46a6-40b6-87b3-b7641b71bf4f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"We also have a fourth criteria which is primarily applicable at higher height","work_id":"16e36ec9-96ac-4895-be8f-1593c9629814","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":12,"snapshot_sha256":"2106365b4047f74aef9e8b24abb976190c515553ae0de5fc813f998be3caf7e5","internal_anchors":0},"formal_canon":{"evidence_count":3,"snapshot_sha256":"20dab4a935667ae522f13712112998461eb3e92f0ee4fd34c6aa5a220b887a4b"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"4593b665-e34c-45c1-8e5b-878911626631"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:13Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"XpFOh/D5CaLIOt6sjhHgRWBVCHUq5ER2byXIGba4ZergWTCLauIJOpHI5VfP6+jHCRv8ECbkrJv6R5trJ9dcDQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-23T19:10:45.575282Z"},"content_sha256":"663ffbfba3c47f49d093b32c8d8bdc5239b343415affdb63da49b806f1b4271b","schema_version":"1.0","event_id":"sha256:663ffbfba3c47f49d093b32c8d8bdc5239b343415affdb63da49b806f1b4271b"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/UKSLWPQT2Z3UNZ3JUCL37JYPQF/bundle.json","state_url":"https://pith.science/pith/UKSLWPQT2Z3UNZ3JUCL37JYPQF/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/UKSLWPQT2Z3UNZ3JUCL37JYPQF/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-23T19:10:45Z","links":{"resolver":"https://pith.science/pith/UKSLWPQT2Z3UNZ3JUCL37JYPQF","bundle":"https://pith.science/pith/UKSLWPQT2Z3UNZ3JUCL37JYPQF/bundle.json","state":"https://pith.science/pith/UKSLWPQT2Z3UNZ3JUCL37JYPQF/state.json","well_known_bundle":"https://pith.science/.well-known/pith/UKSLWPQT2Z3UNZ3JUCL37JYPQF/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2021:UKSLWPQT2Z3UNZ3JUCL37JYPQF","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"ecaab0f0204b68cab469b5a1be1cb0e5693ea4b755132f7c575933fc96022bdb","cross_cats_sorted":["cs.AI","cs.LG"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2021-09-22T17:34:18Z","title_canon_sha256":"612ddce1253fc4e6a3ead618e152978b1fbe6ff31ad268ab30b5b82b9e65fc6e"},"schema_version":"1.0","source":{"id":"2109.10862","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2109.10862","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"arxiv_version","alias_value":"2109.10862v2","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2109.10862","created_at":"2026-05-17T23:38:13Z"},{"alias_kind":"pith_short_12","alias_value":"UKSLWPQT2Z3U","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"UKSLWPQT2Z3UNZ3J","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"UKSLWPQT","created_at":"2026-05-18T12:33:33Z"}],"graph_snapshots":[{"event_id":"sha256:663ffbfba3c47f49d093b32c8d8bdc5239b343415affdb63da49b806f1b4271b","target":"graph","created_at":"2026-05-17T23:38:13Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"Our resulting model generates sensible summaries of entire books, even matching the quality of human-written summaries in a few cases (~5% of books). We achieve state-of-the-art results on the recent BookSum dataset for book-length summarization."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That summaries of summaries retain enough information and fidelity for the final output to remain faithful to the original book when humans never see the full text."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Recursive decomposition plus human feedback lets GPT-3 produce book-length summaries that reach human quality on a few cases and set new records on BookSum and NarrativeQA."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Recursive decomposition lets models summarize entire books after humans give feedback only on short sections."}],"snapshot_sha256":"ac2d3692783a95b8708e95936c8d8684c6c8e48ee69de27882d9644785e03bea"},"formal_canon":{"evidence_count":3,"snapshot_sha256":"20dab4a935667ae522f13712112998461eb3e92f0ee4fd34c6aa5a220b887a4b"},"paper":{"abstract_excerpt":"A major challenge for scaling machine learning is training models to perform tasks that are very difficult or time-consuming for humans to evaluate. We present progress on this problem on the task of abstractive summarization of entire fiction novels. Our method combines learning from human feedback with recursive task decomposition: we use models trained on smaller parts of the task to assist humans in giving feedback on the broader task. We collect a large volume of demonstrations and comparisons from human labelers, and fine-tune GPT-3 using behavioral cloning and reward modeling to do summ","authors_text":"Daniel M. Ziegler, Jan Leike, Jeff Wu, Long Ouyang, Nisan Stiennon, Paul Christiano, Ryan Lowe","cross_cats":["cs.AI","cs.LG"],"headline":"Recursive decomposition lets models summarize entire books after humans give feedback only on short sections.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2021-09-22T17:34:18Z","title":"Recursively Summarizing Books with Human Feedback"},"references":{"count":12,"internal_anchors":0,"resolved_work":12,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"This subtask can be decomposed even further if necessary","work_id":"0edfbdfc-d42f-42ee-85ca-86cabf1cf023","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"answer_directly, which returns an actual answer to the task, synthesizing the answers to subtasks In general, both decompose_if_needed and answer_directly could be learned and implemented by an ML mod","work_id":"a86a2eb4-2be7-457f-b24e-1f5157bb1f02","year":2020},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"So gratuitously including small details is generally penalized, and omitting important details is also penalized","work_id":"18f5ae1f-6b0b-4c05-a46f-7f852267c26b","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Accuracy: All information in the summary should faithfully reflect the original passage","work_id":"0c69f10e-46a6-40b6-87b3-b7641b71bf4f","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"We also have a fourth criteria which is primarily applicable at higher height","work_id":"16e36ec9-96ac-4895-be8f-1593c9629814","year":null}],"snapshot_sha256":"2106365b4047f74aef9e8b24abb976190c515553ae0de5fc813f998be3caf7e5"},"source":{"id":"2109.10862","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-17T15:22:19.704135Z","id":"4593b665-e34c-45c1-8e5b-878911626631","model_set":{"reader":"grok-4.3"},"one_line_summary":"Recursive decomposition plus human feedback lets GPT-3 produce book-length summaries that reach human quality on a few cases and set new records on BookSum and NarrativeQA.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Recursive decomposition lets models summarize entire books after humans give feedback only on short sections.","strongest_claim":"Our resulting model generates sensible summaries of entire books, even matching the quality of human-written summaries in a few cases (~5% of books). We achieve state-of-the-art results on the recent BookSum dataset for book-length summarization.","weakest_assumption":"That summaries of summaries retain enough information and fidelity for the final output to remain faithful to the original book when humans never see the full text."}},"verdict_id":"4593b665-e34c-45c1-8e5b-878911626631"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:de5808fbf6703c24c07c86c3bfff303c374f0c66c8d616868b39aad5d7b0f36a","target":"record","created_at":"2026-05-17T23:38:13Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"ecaab0f0204b68cab469b5a1be1cb0e5693ea4b755132f7c575933fc96022bdb","cross_cats_sorted":["cs.AI","cs.LG"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2021-09-22T17:34:18Z","title_canon_sha256":"612ddce1253fc4e6a3ead618e152978b1fbe6ff31ad268ab30b5b82b9e65fc6e"},"schema_version":"1.0","source":{"id":"2109.10862","kind":"arxiv","version":2}},"canonical_sha256":"a2a4bb3e13d67746e769a097bfa70f8148c475fb3b8ffa86c331751c38e0255a","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"a2a4bb3e13d67746e769a097bfa70f8148c475fb3b8ffa86c331751c38e0255a","first_computed_at":"2026-05-17T23:38:13.696569Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:13.696569Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"5n+v4tOAmms5ajGNA+c42Iub3gqC6Ixo9pa1IxIQ5GTkJJ2wJqbEUKJxZxHzfgfWjevCCzfZ81R5Zg015HEwDg==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:13.697225Z","signed_message":"canonical_sha256_bytes"},"source_id":"2109.10862","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:de5808fbf6703c24c07c86c3bfff303c374f0c66c8d616868b39aad5d7b0f36a","sha256:663ffbfba3c47f49d093b32c8d8bdc5239b343415affdb63da49b806f1b4271b"],"state_sha256":"b3aa435f21053a936cf684a16eaca2ecb76ddb7a77cd1f94eb08bd61acf46b44"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"DpKunv4VBC11ycaTZ7AFkc+RSTnGUYaX3j6L/wDNUlW9DNjJVSggv/3TpOhmsDrf/FfR89ujA5VEaWu8iMqYBg==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-23T19:10:45.578791Z","bundle_sha256":"b14bf9b4adfc9241df35c64d21f19e691912fe0047bb918a5f7565767b5a6ab7"}}