{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:XRUI6OAW7LFJP4QLIS5FVOHNOP","short_pith_number":"pith:XRUI6OAW","schema_version":"1.0","canonical_sha256":"bc688f3816faca97f20b44ba5ab8ed73c41313872d030185bab88a266d9d8061","source":{"kind":"arxiv","id":"2606.21090","version":1},"attestation_state":"computed","paper":{"title":"Self-Improvement Can Self-Regress: The Rise-and-Collapse Failure Mode of LLM Self-Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Jianzhe Lin","submitted_at":"2026-06-17T18:03:06Z","abstract_excerpt":"Self-improvement can self-regress. In REINFORCE post-training for code, a model can quickly improve on its optimized metric and then collapse within the same training campaign. We study this in a controlled multi-seed testbed using Qwen-2.5-3B and Qwen-2.5-7B, trained on competitive-programming tasks with binary CodeGrader reward across 10 sequential 20-step campaigns. Across campaigns, pass@1 shows a robust rise-then-collapse pattern: it peaks within tens of gradient steps and then falls back, sometimes to near zero. This is not cross-task catastrophic forgetting, but within-task policy over-"},"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":"2606.21090","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2026-06-17T18:03:06Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"95b8fc03068143c614557cc86ff7884ad4d403f518f9eecde045bbd4a8bcf08b","abstract_canon_sha256":"79b8df4eaac28179cd550e64db1a683a2d9a9b5c132fe5df349149b15f81a202"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-23T01:12:29.544507Z","signature_b64":"SKq4aqQoWR68fPVNUc/42Y/t8wBtvYpKxhUzV8v3xrtpU/mnO5hCb6MOqWIZfyC+Q/p/54rT0lZ52aPso/VOAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"bc688f3816faca97f20b44ba5ab8ed73c41313872d030185bab88a266d9d8061","last_reissued_at":"2026-06-23T01:12:29.544030Z","signature_status":"signed_v1","first_computed_at":"2026-06-23T01:12:29.544030Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Self-Improvement Can Self-Regress: The Rise-and-Collapse Failure Mode of LLM Self-Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Jianzhe Lin","submitted_at":"2026-06-17T18:03:06Z","abstract_excerpt":"Self-improvement can self-regress. In REINFORCE post-training for code, a model can quickly improve on its optimized metric and then collapse within the same training campaign. We study this in a controlled multi-seed testbed using Qwen-2.5-3B and Qwen-2.5-7B, trained on competitive-programming tasks with binary CodeGrader reward across 10 sequential 20-step campaigns. Across campaigns, pass@1 shows a robust rise-then-collapse pattern: it peaks within tens of gradient steps and then falls back, sometimes to near zero. This is not cross-task catastrophic forgetting, but within-task policy over-"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.21090","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/2606.21090/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":"2606.21090","created_at":"2026-06-23T01:12:29.544097+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.21090v1","created_at":"2026-06-23T01:12:29.544097+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.21090","created_at":"2026-06-23T01:12:29.544097+00:00"},{"alias_kind":"pith_short_12","alias_value":"XRUI6OAW7LFJ","created_at":"2026-06-23T01:12:29.544097+00:00"},{"alias_kind":"pith_short_16","alias_value":"XRUI6OAW7LFJP4QL","created_at":"2026-06-23T01:12:29.544097+00:00"},{"alias_kind":"pith_short_8","alias_value":"XRUI6OAW","created_at":"2026-06-23T01:12:29.544097+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/XRUI6OAW7LFJP4QLIS5FVOHNOP","json":"https://pith.science/pith/XRUI6OAW7LFJP4QLIS5FVOHNOP.json","graph_json":"https://pith.science/api/pith-number/XRUI6OAW7LFJP4QLIS5FVOHNOP/graph.json","events_json":"https://pith.science/api/pith-number/XRUI6OAW7LFJP4QLIS5FVOHNOP/events.json","paper":"https://pith.science/paper/XRUI6OAW"},"agent_actions":{"view_html":"https://pith.science/pith/XRUI6OAW7LFJP4QLIS5FVOHNOP","download_json":"https://pith.science/pith/XRUI6OAW7LFJP4QLIS5FVOHNOP.json","view_paper":"https://pith.science/paper/XRUI6OAW","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.21090&json=true","fetch_graph":"https://pith.science/api/pith-number/XRUI6OAW7LFJP4QLIS5FVOHNOP/graph.json","fetch_events":"https://pith.science/api/pith-number/XRUI6OAW7LFJP4QLIS5FVOHNOP/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XRUI6OAW7LFJP4QLIS5FVOHNOP/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XRUI6OAW7LFJP4QLIS5FVOHNOP/action/storage_attestation","attest_author":"https://pith.science/pith/XRUI6OAW7LFJP4QLIS5FVOHNOP/action/author_attestation","sign_citation":"https://pith.science/pith/XRUI6OAW7LFJP4QLIS5FVOHNOP/action/citation_signature","submit_replication":"https://pith.science/pith/XRUI6OAW7LFJP4QLIS5FVOHNOP/action/replication_record"}},"created_at":"2026-06-23T01:12:29.544097+00:00","updated_at":"2026-06-23T01:12:29.544097+00:00"}