{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:WWKRFPTL57FUW6SVYDP5XPI4TV","short_pith_number":"pith:WWKRFPTL","schema_version":"1.0","canonical_sha256":"b59512be6befcb4b7a55c0dfdbbd1c9d71d6dbd398dad06432d45caab14b4327","source":{"kind":"arxiv","id":"2605.14588","version":1},"attestation_state":"computed","paper":{"title":"Silent Collapse in Recursive Learning Systems","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Recursive models lose internal diversity even as standard metrics remain stable or improve.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Zhipeng Zhang","submitted_at":"2026-05-14T08:59:26Z","abstract_excerpt":"Recursive learning -- where models are trained on data generated by previous versions of themselves -- is increasingly common in large language models, autonomous agents, and self-supervised systems. However, standard performance metrics (loss, perplexity, accuracy) often fail to detect internal degradation before it becomes irreversible. Here we identify a phenomenon we call silent collapse: under broad recursive conditions, model internal distributions -- predictive entropy, representational diversity, and tail coverage -- progressively contract even as conventional metrics appear stable or "},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.14588","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-14T08:59:26Z","cross_cats_sorted":[],"title_canon_sha256":"191797cb5e6fe65f4862420a0ef09e0742dddc66f6caac4888e6aff0ff04fa78","abstract_canon_sha256":"5db07391d10b542455d1e754ac6ac323d33a0a3b89f5ed536b2612e8e6d786a9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:05.279569Z","signature_b64":"glojHuYH7FP9HxVXRLug8f15XSalmk3ZeLMvfr/6MhNOBqkwgEjzcFe+ArrGl3AtyNLbQ65gpjI7loUjRknaBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b59512be6befcb4b7a55c0dfdbbd1c9d71d6dbd398dad06432d45caab14b4327","last_reissued_at":"2026-05-17T23:39:05.278956Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:05.278956Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Silent Collapse in Recursive Learning Systems","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Recursive models lose internal diversity even as standard metrics remain stable or improve.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Zhipeng Zhang","submitted_at":"2026-05-14T08:59:26Z","abstract_excerpt":"Recursive learning -- where models are trained on data generated by previous versions of themselves -- is increasingly common in large language models, autonomous agents, and self-supervised systems. However, standard performance metrics (loss, perplexity, accuracy) often fail to detect internal degradation before it becomes irreversible. Here we identify a phenomenon we call silent collapse: under broad recursive conditions, model internal distributions -- predictive entropy, representational diversity, and tail coverage -- progressively contract even as conventional metrics appear stable or "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"under broad recursive conditions, model internal distributions -- predictive entropy, representational diversity, and tail coverage -- progressively contract even as conventional metrics appear stable or improving.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The three trajectory-level precursors reliably appear multiple generations before any degradation in standard validation metrics, and the MTR loop can estimate trust and modulate learning intensity effectively without access to pristine real data.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Recursive learning systems undergo silent collapse of internal distributions, preceded by entropy contraction, representation freezing, and tail erosion, which the MTR framework can monitor and avert.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Recursive models lose internal diversity even as standard metrics remain stable or improve.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6b4a8c4d7568ba1317985ef30b10880a489edd45056abe1265f2a2ea3390848b"},"source":{"id":"2605.14588","kind":"arxiv","version":1},"verdict":{"id":"019e0a0e-27dc-499b-a707-6cff9bfdfc90","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:27:13.450953Z","strongest_claim":"under broad recursive conditions, model internal distributions -- predictive entropy, representational diversity, and tail coverage -- progressively contract even as conventional metrics appear stable or improving.","one_line_summary":"Recursive learning systems undergo silent collapse of internal distributions, preceded by entropy contraction, representation freezing, and tail erosion, which the MTR framework can monitor and avert.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The three trajectory-level precursors reliably appear multiple generations before any degradation in standard validation metrics, and the MTR loop can estimate trust and modulate learning intensity effectively without access to pristine real data.","pith_extraction_headline":"Recursive models lose internal diversity even as standard metrics remain stable or improve."},"references":{"count":16,"sample":[{"doi":"","year":2019,"title":"Language models are unsupervised multitask learners.OpenAI Blog, 1(8):9, 2019","work_id":"bc2954ac-13c2-4e17-a960-1c62ce7d2aea","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1901,"title":"Brown, Benjamin Mann, Nick Ryder, et al","work_id":"6f0d940a-3b3f-4940-8013-c53aee6da3e4","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Christiano, Jan Leike, Tom B","work_id":"c3896b8b-9bc0-48c8-b4d9-fcf445c25727","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"Concrete problems in ai safety","work_id":"b8af615c-dbfc-43cc-8db1-38d3eeca49c3","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Ai models collapse when trained on recursively generated data.Nature, 632:755–760,","work_id":"3c9e452b-b160-4949-8280-30174ddc2155","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":16,"snapshot_sha256":"a4d98b2f825d0d8ec93886dd25f40bc378c4e84ad51361d01cdec5c33b62bc92","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"f2147c21a1068a23b415ef86965dfeb2d059d7b366cc1630ec48ea207adad6b6"},"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":"2605.14588","created_at":"2026-05-17T23:39:05.279057+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.14588v1","created_at":"2026-05-17T23:39:05.279057+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14588","created_at":"2026-05-17T23:39:05.279057+00:00"},{"alias_kind":"pith_short_12","alias_value":"WWKRFPTL57FU","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"WWKRFPTL57FUW6SV","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"WWKRFPTL","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/WWKRFPTL57FUW6SVYDP5XPI4TV","json":"https://pith.science/pith/WWKRFPTL57FUW6SVYDP5XPI4TV.json","graph_json":"https://pith.science/api/pith-number/WWKRFPTL57FUW6SVYDP5XPI4TV/graph.json","events_json":"https://pith.science/api/pith-number/WWKRFPTL57FUW6SVYDP5XPI4TV/events.json","paper":"https://pith.science/paper/WWKRFPTL"},"agent_actions":{"view_html":"https://pith.science/pith/WWKRFPTL57FUW6SVYDP5XPI4TV","download_json":"https://pith.science/pith/WWKRFPTL57FUW6SVYDP5XPI4TV.json","view_paper":"https://pith.science/paper/WWKRFPTL","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.14588&json=true","fetch_graph":"https://pith.science/api/pith-number/WWKRFPTL57FUW6SVYDP5XPI4TV/graph.json","fetch_events":"https://pith.science/api/pith-number/WWKRFPTL57FUW6SVYDP5XPI4TV/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/WWKRFPTL57FUW6SVYDP5XPI4TV/action/timestamp_anchor","attest_storage":"https://pith.science/pith/WWKRFPTL57FUW6SVYDP5XPI4TV/action/storage_attestation","attest_author":"https://pith.science/pith/WWKRFPTL57FUW6SVYDP5XPI4TV/action/author_attestation","sign_citation":"https://pith.science/pith/WWKRFPTL57FUW6SVYDP5XPI4TV/action/citation_signature","submit_replication":"https://pith.science/pith/WWKRFPTL57FUW6SVYDP5XPI4TV/action/replication_record"}},"created_at":"2026-05-17T23:39:05.279057+00:00","updated_at":"2026-05-17T23:39:05.279057+00:00"}