{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:7SFZPTLFZDVY4NVBUGZSLTPHAY","short_pith_number":"pith:7SFZPTLF","schema_version":"1.0","canonical_sha256":"fc8b97cd65c8eb8e36a1a1b325cde7061f20029932b41787cfcd40b392ffc9de","source":{"kind":"arxiv","id":"2607.27240","version":1},"attestation_state":"computed","paper":{"title":"Asymmetric Collapse in Model Merging: When Refusal Over- writes Recognition","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Aarnav Choudhary, Jiwon Seo, Maheep Chaudhary, Matheus Fonseca Rocha, Vasu Sharma","submitted_at":"2026-07-26T19:10:34Z","abstract_excerpt":"Model merging is often used to combine capabilities from separately fine-tuned models without additional training, but it is unclear whether standard merging methods preserve multiple safety-relevant behaviors simultaneously. We study this question through a controlled case study using two Gemma-3-1B-IT finetunes on two complementary safety objectives: CARES harm-level classification and WildJailbreak adversarial refusal. We merge the two fine-tunes using Linear, SLERP, TIES, and DARE-TIES, and evaluate the merged models on classification accuracy, attack resistance, and benign compliance. Acr"},"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":"2607.27240","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-07-26T19:10:34Z","cross_cats_sorted":[],"title_canon_sha256":"70742c8f4be033a38615071c5c917f881841f4ac7db01ed04c1be6ebd5d30800","abstract_canon_sha256":"6ffd69b9bdd13fc2eaf1054bcca57875f5865d2df6dcb9e9a0ede5e178a3b78b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"fc8b97cd65c8eb8e36a1a1b325cde7061f20029932b41787cfcd40b392ffc9de","last_reissued_at":"2026-07-31T00:10:23.753213Z","signature_status":"unsigned_v0","first_computed_at":"2026-07-31T00:10:23.753213Z"},"graph_snapshot":{"paper":{"title":"Asymmetric Collapse in Model Merging: When Refusal Over- writes Recognition","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Aarnav Choudhary, Jiwon Seo, Maheep Chaudhary, Matheus Fonseca Rocha, Vasu Sharma","submitted_at":"2026-07-26T19:10:34Z","abstract_excerpt":"Model merging is often used to combine capabilities from separately fine-tuned models without additional training, but it is unclear whether standard merging methods preserve multiple safety-relevant behaviors simultaneously. We study this question through a controlled case study using two Gemma-3-1B-IT finetunes on two complementary safety objectives: CARES harm-level classification and WildJailbreak adversarial refusal. We merge the two fine-tunes using Linear, SLERP, TIES, and DARE-TIES, and evaluate the merged models on classification accuracy, attack resistance, and benign compliance. Acr"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2607.27240","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/2607.27240/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":"2607.27240","created_at":"2026-07-31T00:10:23.755584+00:00"},{"alias_kind":"arxiv_version","alias_value":"2607.27240v1","created_at":"2026-07-31T00:10:23.755584+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2607.27240","created_at":"2026-07-31T00:10:23.755584+00:00"},{"alias_kind":"pith_short_12","alias_value":"7SFZPTLFZDVY","created_at":"2026-07-31T00:10:23.755584+00:00"},{"alias_kind":"pith_short_16","alias_value":"7SFZPTLFZDVY4NVB","created_at":"2026-07-31T00:10:23.755584+00:00"},{"alias_kind":"pith_short_8","alias_value":"7SFZPTLF","created_at":"2026-07-31T00:10:23.755584+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/7SFZPTLFZDVY4NVBUGZSLTPHAY","json":"https://pith.science/pith/7SFZPTLFZDVY4NVBUGZSLTPHAY.json","graph_json":"https://pith.science/api/pith-number/7SFZPTLFZDVY4NVBUGZSLTPHAY/graph.json","events_json":"https://pith.science/api/pith-number/7SFZPTLFZDVY4NVBUGZSLTPHAY/events.json","paper":"https://pith.science/paper/7SFZPTLF"},"agent_actions":{"view_html":"https://pith.science/pith/7SFZPTLFZDVY4NVBUGZSLTPHAY","download_json":"https://pith.science/pith/7SFZPTLFZDVY4NVBUGZSLTPHAY.json","view_paper":"https://pith.science/paper/7SFZPTLF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2607.27240&json=true","fetch_graph":"https://pith.science/api/pith-number/7SFZPTLFZDVY4NVBUGZSLTPHAY/graph.json","fetch_events":"https://pith.science/api/pith-number/7SFZPTLFZDVY4NVBUGZSLTPHAY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/7SFZPTLFZDVY4NVBUGZSLTPHAY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/7SFZPTLFZDVY4NVBUGZSLTPHAY/action/storage_attestation","attest_author":"https://pith.science/pith/7SFZPTLFZDVY4NVBUGZSLTPHAY/action/author_attestation","sign_citation":"https://pith.science/pith/7SFZPTLFZDVY4NVBUGZSLTPHAY/action/citation_signature","submit_replication":"https://pith.science/pith/7SFZPTLFZDVY4NVBUGZSLTPHAY/action/replication_record"}},"created_at":"2026-07-31T00:10:23.755584+00:00","updated_at":"2026-07-31T00:10:23.755584+00:00"}