{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:IJBMPAA4CWIQGWPJ4MANEG4XJJ","short_pith_number":"pith:IJBMPAA4","schema_version":"1.0","canonical_sha256":"4242c7801c15910359e9e300d21b974a529cd048dbb005d5eab962dc9bf7afef","source":{"kind":"arxiv","id":"2509.04796","version":1},"attestation_state":"computed","paper":{"title":"Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Adriano Koshiyama, Figarri Keisha, Philip Treleaven, Zekun Wu, Ze Wang","submitted_at":"2025-09-05T04:29:15Z","abstract_excerpt":"Large language models increasingly rely on synthetic data due to human-written content scarcity, yet recursive training on model-generated outputs leads to model collapse, a degenerative process threatening factual reliability. We define knowledge collapse as a distinct three-stage phenomenon where factual accuracy deteriorates while surface fluency persists, creating \"confidently wrong\" outputs that pose critical risks in accuracy-dependent domains. Through controlled experiments with recursive synthetic training, we demonstrate that collapse trajectory and timing depend critically on instruc"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2509.04796","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.CL","submitted_at":"2025-09-05T04:29:15Z","cross_cats_sorted":[],"title_canon_sha256":"7ec8796e2a9407fafdc1f9984bb8d9f5c62612c3b76521d77442c6f2788eef62","abstract_canon_sha256":"27c1f3978ccbab81c4f1227605a8af2a6ebb38496a280e64d75674f4ff1aef7b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T12:05:29.859942Z","signature_b64":"Yo9GT3QiaMx+7TX4q1ezl9CByULi3Qeyix52HiRFM77MKanrwf7/6Ay7K104GnAhJ/J/rto2xo3fuyVx71WoBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4242c7801c15910359e9e300d21b974a529cd048dbb005d5eab962dc9bf7afef","last_reissued_at":"2026-07-05T12:05:29.859346Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T12:05:29.859346Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Adriano Koshiyama, Figarri Keisha, Philip Treleaven, Zekun Wu, Ze Wang","submitted_at":"2025-09-05T04:29:15Z","abstract_excerpt":"Large language models increasingly rely on synthetic data due to human-written content scarcity, yet recursive training on model-generated outputs leads to model collapse, a degenerative process threatening factual reliability. We define knowledge collapse as a distinct three-stage phenomenon where factual accuracy deteriorates while surface fluency persists, creating \"confidently wrong\" outputs that pose critical risks in accuracy-dependent domains. Through controlled experiments with recursive synthetic training, we demonstrate that collapse trajectory and timing depend critically on instruc"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2509.04796","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2509.04796/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2509.04796","created_at":"2026-07-05T12:05:29.859427+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.04796v1","created_at":"2026-07-05T12:05:29.859427+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.04796","created_at":"2026-07-05T12:05:29.859427+00:00"},{"alias_kind":"pith_short_12","alias_value":"IJBMPAA4CWIQ","created_at":"2026-07-05T12:05:29.859427+00:00"},{"alias_kind":"pith_short_16","alias_value":"IJBMPAA4CWIQGWPJ","created_at":"2026-07-05T12:05:29.859427+00:00"},{"alias_kind":"pith_short_8","alias_value":"IJBMPAA4","created_at":"2026-07-05T12:05:29.859427+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.20602","citing_title":"Self-Training Doesn't Flatten Language -- It Restructures It: Surface Markers Amplify While Deep Syntax Dies","ref_index":20,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/IJBMPAA4CWIQGWPJ4MANEG4XJJ","json":"https://pith.science/pith/IJBMPAA4CWIQGWPJ4MANEG4XJJ.json","graph_json":"https://pith.science/api/pith-number/IJBMPAA4CWIQGWPJ4MANEG4XJJ/graph.json","events_json":"https://pith.science/api/pith-number/IJBMPAA4CWIQGWPJ4MANEG4XJJ/events.json","paper":"https://pith.science/paper/IJBMPAA4"},"agent_actions":{"view_html":"https://pith.science/pith/IJBMPAA4CWIQGWPJ4MANEG4XJJ","download_json":"https://pith.science/pith/IJBMPAA4CWIQGWPJ4MANEG4XJJ.json","view_paper":"https://pith.science/paper/IJBMPAA4","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.04796&json=true","fetch_graph":"https://pith.science/api/pith-number/IJBMPAA4CWIQGWPJ4MANEG4XJJ/graph.json","fetch_events":"https://pith.science/api/pith-number/IJBMPAA4CWIQGWPJ4MANEG4XJJ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/IJBMPAA4CWIQGWPJ4MANEG4XJJ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/IJBMPAA4CWIQGWPJ4MANEG4XJJ/action/storage_attestation","attest_author":"https://pith.science/pith/IJBMPAA4CWIQGWPJ4MANEG4XJJ/action/author_attestation","sign_citation":"https://pith.science/pith/IJBMPAA4CWIQGWPJ4MANEG4XJJ/action/citation_signature","submit_replication":"https://pith.science/pith/IJBMPAA4CWIQGWPJ4MANEG4XJJ/action/replication_record"}},"created_at":"2026-07-05T12:05:29.859427+00:00","updated_at":"2026-07-05T12:05:29.859427+00:00"}