{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:TNUXS2DUYAPM5BB2ZSGJM7TFFV","short_pith_number":"pith:TNUXS2DU","schema_version":"1.0","canonical_sha256":"9b69796874c01ece843acc8c967e652d66bdb358abe32d0c8dadab13dff69f4e","source":{"kind":"arxiv","id":"2602.07340","version":2},"attestation_state":"computed","paper":{"title":"Revisiting Robustness for LLM Safety Alignment via Selective Geometry Control","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jilong Liu, Junfeng Fang, Le Wu, Richang Hong, Tat-Sent Chua, Weibiao Huang, Wenjian Tao, Xingyu Zhu, Yonghui Yang","submitted_at":"2026-02-07T03:46:33Z","abstract_excerpt":"Safety alignment of large language models remains brittle under domain shift and noisy preference supervision. Most existing robust alignment methods focus on uncertainty in alignment data, while overlooking optimization-induced fragility in preference-based objectives. In this work, we revisit robustness for LLM safety alignment from an optimization geometry perspective, and argue that robustness failures cannot be addressed by data-centric methods alone. We propose \\textit{ShaPO}, a geometry-aware preference optimization framework that enforces worst-case alignment objectives via selective g"},"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":"2602.07340","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-02-07T03:46:33Z","cross_cats_sorted":[],"title_canon_sha256":"1ba0a94ca829f1f62c93edf46814b209b73f06eeea7e8da6ff9d6cd03579ae44","abstract_canon_sha256":"1aca1e98e1946096cfe0fcf7c450251f06274a357adffa7ab53d41b3c4808bd1"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-22T01:03:56.856527Z","signature_b64":"qo0lAq9v5dAJS8YIwkNDxzEjX0iqmnsqkDhGjb6Mi7NpOEZcxdSdCRjUpZVQDiTBbwu7Vgqh5RQRwl4aWqBgBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9b69796874c01ece843acc8c967e652d66bdb358abe32d0c8dadab13dff69f4e","last_reissued_at":"2026-05-22T01:03:56.855978Z","signature_status":"signed_v1","first_computed_at":"2026-05-22T01:03:56.855978Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Revisiting Robustness for LLM Safety Alignment via Selective Geometry Control","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jilong Liu, Junfeng Fang, Le Wu, Richang Hong, Tat-Sent Chua, Weibiao Huang, Wenjian Tao, Xingyu Zhu, Yonghui Yang","submitted_at":"2026-02-07T03:46:33Z","abstract_excerpt":"Safety alignment of large language models remains brittle under domain shift and noisy preference supervision. Most existing robust alignment methods focus on uncertainty in alignment data, while overlooking optimization-induced fragility in preference-based objectives. In this work, we revisit robustness for LLM safety alignment from an optimization geometry perspective, and argue that robustness failures cannot be addressed by data-centric methods alone. We propose \\textit{ShaPO}, a geometry-aware preference optimization framework that enforces worst-case alignment objectives via selective g"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.07340","kind":"arxiv","version":2},"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/2602.07340/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":"2602.07340","created_at":"2026-05-22T01:03:56.856040+00:00"},{"alias_kind":"arxiv_version","alias_value":"2602.07340v2","created_at":"2026-05-22T01:03:56.856040+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.07340","created_at":"2026-05-22T01:03:56.856040+00:00"},{"alias_kind":"pith_short_12","alias_value":"TNUXS2DUYAPM","created_at":"2026-05-22T01:03:56.856040+00:00"},{"alias_kind":"pith_short_16","alias_value":"TNUXS2DUYAPM5BB2","created_at":"2026-05-22T01:03:56.856040+00:00"},{"alias_kind":"pith_short_8","alias_value":"TNUXS2DU","created_at":"2026-05-22T01:03:56.856040+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/TNUXS2DUYAPM5BB2ZSGJM7TFFV","json":"https://pith.science/pith/TNUXS2DUYAPM5BB2ZSGJM7TFFV.json","graph_json":"https://pith.science/api/pith-number/TNUXS2DUYAPM5BB2ZSGJM7TFFV/graph.json","events_json":"https://pith.science/api/pith-number/TNUXS2DUYAPM5BB2ZSGJM7TFFV/events.json","paper":"https://pith.science/paper/TNUXS2DU"},"agent_actions":{"view_html":"https://pith.science/pith/TNUXS2DUYAPM5BB2ZSGJM7TFFV","download_json":"https://pith.science/pith/TNUXS2DUYAPM5BB2ZSGJM7TFFV.json","view_paper":"https://pith.science/paper/TNUXS2DU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2602.07340&json=true","fetch_graph":"https://pith.science/api/pith-number/TNUXS2DUYAPM5BB2ZSGJM7TFFV/graph.json","fetch_events":"https://pith.science/api/pith-number/TNUXS2DUYAPM5BB2ZSGJM7TFFV/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/TNUXS2DUYAPM5BB2ZSGJM7TFFV/action/timestamp_anchor","attest_storage":"https://pith.science/pith/TNUXS2DUYAPM5BB2ZSGJM7TFFV/action/storage_attestation","attest_author":"https://pith.science/pith/TNUXS2DUYAPM5BB2ZSGJM7TFFV/action/author_attestation","sign_citation":"https://pith.science/pith/TNUXS2DUYAPM5BB2ZSGJM7TFFV/action/citation_signature","submit_replication":"https://pith.science/pith/TNUXS2DUYAPM5BB2ZSGJM7TFFV/action/replication_record"}},"created_at":"2026-05-22T01:03:56.856040+00:00","updated_at":"2026-05-22T01:03:56.856040+00:00"}