{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:OTXUMK3OKIN3LKFYTYUNF5KUKR","short_pith_number":"pith:OTXUMK3O","schema_version":"1.0","canonical_sha256":"74ef462b6e521bb5a8b89e28d2f554545099e7d00d0fbec73ef34fbd0f37d9e1","source":{"kind":"arxiv","id":"2605.28860","version":1},"attestation_state":"computed","paper":{"title":"Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Jeanmely Rojas Nunez, Maheep Chaudhary, Nathan Allen, Nomgondalai Amgalanbaatar, Vasu Sharma, Viraj Sawant, Yannis Zongo","submitted_at":"2026-05-21T19:03:57Z","abstract_excerpt":"Fine-tuning large language models (LLMs) frequently induces catastrophic forgetting of prior capabilities. Recent work has shown that reinforcement learning (RL) retains prior capabilities more effectively than supervised fine-tuning (SFT), attributing this to policy-gradient updates remaining closer to the base policy \\cite{shenfeld2025rl}. We extend this behavioral account to the mechanistic level and ask whether RL's advantage is mirrored by stronger preservation of internal computational circuits. We introduce differential circuit vulnerability, a head-level measure of how much a circuit d"},"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":"2605.28860","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-21T19:03:57Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"dc0eff3b35b99a89a12495c0f7dce7825c09041fe133b78fbc2d5f896aeb975b","abstract_canon_sha256":"391b526d007c29deb649d6db5fdabdb3a207a280b586b6c4ccebe2feddba8f84"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-29T00:04:14.534341Z","signature_b64":"GjNkuiwSNkVUzntDwDg5/qHw5j8RJueTjwtT6PN9feCFRUaW42Bv5u5wQWECbFmaD7KKDErK9k2fgmlIEcFnAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"74ef462b6e521bb5a8b89e28d2f554545099e7d00d0fbec73ef34fbd0f37d9e1","last_reissued_at":"2026-05-29T00:04:14.533794Z","signature_status":"signed_v1","first_computed_at":"2026-05-29T00:04:14.533794Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Jeanmely Rojas Nunez, Maheep Chaudhary, Nathan Allen, Nomgondalai Amgalanbaatar, Vasu Sharma, Viraj Sawant, Yannis Zongo","submitted_at":"2026-05-21T19:03:57Z","abstract_excerpt":"Fine-tuning large language models (LLMs) frequently induces catastrophic forgetting of prior capabilities. Recent work has shown that reinforcement learning (RL) retains prior capabilities more effectively than supervised fine-tuning (SFT), attributing this to policy-gradient updates remaining closer to the base policy \\cite{shenfeld2025rl}. We extend this behavioral account to the mechanistic level and ask whether RL's advantage is mirrored by stronger preservation of internal computational circuits. We introduce differential circuit vulnerability, a head-level measure of how much a circuit d"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.28860","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/2605.28860/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":"2605.28860","created_at":"2026-05-29T00:04:14.533863+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.28860v1","created_at":"2026-05-29T00:04:14.533863+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.28860","created_at":"2026-05-29T00:04:14.533863+00:00"},{"alias_kind":"pith_short_12","alias_value":"OTXUMK3OKIN3","created_at":"2026-05-29T00:04:14.533863+00:00"},{"alias_kind":"pith_short_16","alias_value":"OTXUMK3OKIN3LKFY","created_at":"2026-05-29T00:04:14.533863+00:00"},{"alias_kind":"pith_short_8","alias_value":"OTXUMK3O","created_at":"2026-05-29T00:04:14.533863+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/OTXUMK3OKIN3LKFYTYUNF5KUKR","json":"https://pith.science/pith/OTXUMK3OKIN3LKFYTYUNF5KUKR.json","graph_json":"https://pith.science/api/pith-number/OTXUMK3OKIN3LKFYTYUNF5KUKR/graph.json","events_json":"https://pith.science/api/pith-number/OTXUMK3OKIN3LKFYTYUNF5KUKR/events.json","paper":"https://pith.science/paper/OTXUMK3O"},"agent_actions":{"view_html":"https://pith.science/pith/OTXUMK3OKIN3LKFYTYUNF5KUKR","download_json":"https://pith.science/pith/OTXUMK3OKIN3LKFYTYUNF5KUKR.json","view_paper":"https://pith.science/paper/OTXUMK3O","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.28860&json=true","fetch_graph":"https://pith.science/api/pith-number/OTXUMK3OKIN3LKFYTYUNF5KUKR/graph.json","fetch_events":"https://pith.science/api/pith-number/OTXUMK3OKIN3LKFYTYUNF5KUKR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OTXUMK3OKIN3LKFYTYUNF5KUKR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OTXUMK3OKIN3LKFYTYUNF5KUKR/action/storage_attestation","attest_author":"https://pith.science/pith/OTXUMK3OKIN3LKFYTYUNF5KUKR/action/author_attestation","sign_citation":"https://pith.science/pith/OTXUMK3OKIN3LKFYTYUNF5KUKR/action/citation_signature","submit_replication":"https://pith.science/pith/OTXUMK3OKIN3LKFYTYUNF5KUKR/action/replication_record"}},"created_at":"2026-05-29T00:04:14.533863+00:00","updated_at":"2026-05-29T00:04:14.533863+00:00"}