{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:PWMHEGTUBP4FRAP7JJCWL3ZABI","short_pith_number":"pith:PWMHEGTU","schema_version":"1.0","canonical_sha256":"7d98721a740bf85881ff4a4565ef200a05979dae5fbd643f0571bca49a7f77d2","source":{"kind":"arxiv","id":"2605.20602","version":1},"attestation_state":"computed","paper":{"title":"Self-Training Doesn't Flatten Language -- It Restructures It: Surface Markers Amplify While Deep Syntax Dies","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Ming Liu","submitted_at":"2026-05-20T01:44:47Z","abstract_excerpt":"Successive self-training on a language model's own outputs is widely characterized as a process of flattening: diversity drops, distributions narrow, and the text becomes \"more like itself.\" We provide evidence that this characterization is incomplete. Across eleven generations of self-training on five models (GPT-2 124M, Pythia-410M, Pythia-1.4B, OPT-1.3B, Pythia-2.8B), language is not flattened uniformly -- it is restructured. Surface markers (discourse connectives, hedges, em-dashes) rise, while mid- and deep-syntactic structures (questions, parentheticals, passives, subjunctives) collapse."},"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":"2605.20602","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-20T01:44:47Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"c2b92730954f699c17986f2306f189c4b470b12ee51577df062b2bbfe8917877","abstract_canon_sha256":"42b0487437fa7b483d7e700031afde1dd6fe1082f777a1452429c9abf4871392"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-21T01:04:44.118097Z","signature_b64":"hNfNlMmhNKBicm+2AN79+Aa2FtkPPry4Z8ogGpRKp7t3qLL/oW+Myss8n+MMi0pTtDs9lZJ5OddF/WiNjYrCDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7d98721a740bf85881ff4a4565ef200a05979dae5fbd643f0571bca49a7f77d2","last_reissued_at":"2026-05-21T01:04:44.117490Z","signature_status":"signed_v1","first_computed_at":"2026-05-21T01:04:44.117490Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Self-Training Doesn't Flatten Language -- It Restructures It: Surface Markers Amplify While Deep Syntax Dies","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Ming Liu","submitted_at":"2026-05-20T01:44:47Z","abstract_excerpt":"Successive self-training on a language model's own outputs is widely characterized as a process of flattening: diversity drops, distributions narrow, and the text becomes \"more like itself.\" We provide evidence that this characterization is incomplete. Across eleven generations of self-training on five models (GPT-2 124M, Pythia-410M, Pythia-1.4B, OPT-1.3B, Pythia-2.8B), language is not flattened uniformly -- it is restructured. Surface markers (discourse connectives, hedges, em-dashes) rise, while mid- and deep-syntactic structures (questions, parentheticals, passives, subjunctives) collapse."},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.20602","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/2605.20602/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":"2605.20602","created_at":"2026-05-21T01:04:44.117585+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.20602v1","created_at":"2026-05-21T01:04:44.117585+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.20602","created_at":"2026-05-21T01:04:44.117585+00:00"},{"alias_kind":"pith_short_12","alias_value":"PWMHEGTUBP4F","created_at":"2026-05-21T01:04:44.117585+00:00"},{"alias_kind":"pith_short_16","alias_value":"PWMHEGTUBP4FRAP7","created_at":"2026-05-21T01:04:44.117585+00:00"},{"alias_kind":"pith_short_8","alias_value":"PWMHEGTU","created_at":"2026-05-21T01:04:44.117585+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PWMHEGTUBP4FRAP7JJCWL3ZABI","json":"https://pith.science/pith/PWMHEGTUBP4FRAP7JJCWL3ZABI.json","graph_json":"https://pith.science/api/pith-number/PWMHEGTUBP4FRAP7JJCWL3ZABI/graph.json","events_json":"https://pith.science/api/pith-number/PWMHEGTUBP4FRAP7JJCWL3ZABI/events.json","paper":"https://pith.science/paper/PWMHEGTU"},"agent_actions":{"view_html":"https://pith.science/pith/PWMHEGTUBP4FRAP7JJCWL3ZABI","download_json":"https://pith.science/pith/PWMHEGTUBP4FRAP7JJCWL3ZABI.json","view_paper":"https://pith.science/paper/PWMHEGTU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.20602&json=true","fetch_graph":"https://pith.science/api/pith-number/PWMHEGTUBP4FRAP7JJCWL3ZABI/graph.json","fetch_events":"https://pith.science/api/pith-number/PWMHEGTUBP4FRAP7JJCWL3ZABI/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PWMHEGTUBP4FRAP7JJCWL3ZABI/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PWMHEGTUBP4FRAP7JJCWL3ZABI/action/storage_attestation","attest_author":"https://pith.science/pith/PWMHEGTUBP4FRAP7JJCWL3ZABI/action/author_attestation","sign_citation":"https://pith.science/pith/PWMHEGTUBP4FRAP7JJCWL3ZABI/action/citation_signature","submit_replication":"https://pith.science/pith/PWMHEGTUBP4FRAP7JJCWL3ZABI/action/replication_record"}},"created_at":"2026-05-21T01:04:44.117585+00:00","updated_at":"2026-05-21T01:04:44.117585+00:00"}