{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:6J3DFGYFRAFTOLFHIH7YFU7VCM","short_pith_number":"pith:6J3DFGYF","schema_version":"1.0","canonical_sha256":"f276329b05880b372ca741ff82d3f5130b7a79059eff5ab0213114ede0e85ed0","source":{"kind":"arxiv","id":"2504.09420","version":1},"attestation_state":"computed","paper":{"title":"SaRO: Enhancing LLM Safety through Reasoning-based Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Shikun Zhang, Wei Ye, Yutao Mou, Yuxiao Luo","submitted_at":"2025-04-13T03:36:06Z","abstract_excerpt":"Current safety alignment techniques for large language models (LLMs) face two key challenges: (1) under-generalization, which leaves models vulnerable to novel jailbreak attacks, and (2) over-alignment, which leads to the excessive refusal of benign instructions. Our preliminary investigation reveals semantic overlap between jailbreak/harmful queries and normal prompts in embedding space, suggesting that more effective safety alignment requires a deeper semantic understanding. This motivates us to incorporate safety-policy-driven reasoning into the alignment process. To this end, we propose th"},"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":"2504.09420","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-04-13T03:36:06Z","cross_cats_sorted":[],"title_canon_sha256":"cb2d1a39f6dea5a33805a666a2cfbe8d047229da3e7d99665298cb19c9d5ec40","abstract_canon_sha256":"b4bd2d51feb55f5ebaab40885af83e93db221c0d26d61ed1dcd29b5784fd1099"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:48:36.833324Z","signature_b64":"9cqSCK1L1GUf9/n3Dq5Dtb3tT4ptvhEi5Pvo7eWiFYP2HCvN6HcPkoad9XxLz5zUrgiSjOu0d7StHvv8idSJBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f276329b05880b372ca741ff82d3f5130b7a79059eff5ab0213114ede0e85ed0","last_reissued_at":"2026-07-05T10:48:36.832872Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:48:36.832872Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SaRO: Enhancing LLM Safety through Reasoning-based Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Shikun Zhang, Wei Ye, Yutao Mou, Yuxiao Luo","submitted_at":"2025-04-13T03:36:06Z","abstract_excerpt":"Current safety alignment techniques for large language models (LLMs) face two key challenges: (1) under-generalization, which leaves models vulnerable to novel jailbreak attacks, and (2) over-alignment, which leads to the excessive refusal of benign instructions. Our preliminary investigation reveals semantic overlap between jailbreak/harmful queries and normal prompts in embedding space, suggesting that more effective safety alignment requires a deeper semantic understanding. This motivates us to incorporate safety-policy-driven reasoning into the alignment process. To this end, we propose th"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.09420","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/2504.09420/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":"2504.09420","created_at":"2026-07-05T10:48:36.832928+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.09420v1","created_at":"2026-07-05T10:48:36.832928+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.09420","created_at":"2026-07-05T10:48:36.832928+00:00"},{"alias_kind":"pith_short_12","alias_value":"6J3DFGYFRAFT","created_at":"2026-07-05T10:48:36.832928+00:00"},{"alias_kind":"pith_short_16","alias_value":"6J3DFGYFRAFTOLFH","created_at":"2026-07-05T10:48:36.832928+00:00"},{"alias_kind":"pith_short_8","alias_value":"6J3DFGYF","created_at":"2026-07-05T10:48:36.832928+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.31748","citing_title":"Addressing Over-Refusal in LLMs with Competing Rewards","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.29396","citing_title":"Aligned but Fragile: Enhancing LLM Safety Robustness via Zeroth-Order Optimization","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2603.17305","citing_title":"Contrastive Reasoning Alignment: Reinforcement Learning from Hidden Representations","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08936","citing_title":"Self-ReSET: Learning to Self-Recover from Unsafe Reasoning Trajectories","ref_index":20,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6J3DFGYFRAFTOLFHIH7YFU7VCM","json":"https://pith.science/pith/6J3DFGYFRAFTOLFHIH7YFU7VCM.json","graph_json":"https://pith.science/api/pith-number/6J3DFGYFRAFTOLFHIH7YFU7VCM/graph.json","events_json":"https://pith.science/api/pith-number/6J3DFGYFRAFTOLFHIH7YFU7VCM/events.json","paper":"https://pith.science/paper/6J3DFGYF"},"agent_actions":{"view_html":"https://pith.science/pith/6J3DFGYFRAFTOLFHIH7YFU7VCM","download_json":"https://pith.science/pith/6J3DFGYFRAFTOLFHIH7YFU7VCM.json","view_paper":"https://pith.science/paper/6J3DFGYF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.09420&json=true","fetch_graph":"https://pith.science/api/pith-number/6J3DFGYFRAFTOLFHIH7YFU7VCM/graph.json","fetch_events":"https://pith.science/api/pith-number/6J3DFGYFRAFTOLFHIH7YFU7VCM/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6J3DFGYFRAFTOLFHIH7YFU7VCM/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6J3DFGYFRAFTOLFHIH7YFU7VCM/action/storage_attestation","attest_author":"https://pith.science/pith/6J3DFGYFRAFTOLFHIH7YFU7VCM/action/author_attestation","sign_citation":"https://pith.science/pith/6J3DFGYFRAFTOLFHIH7YFU7VCM/action/citation_signature","submit_replication":"https://pith.science/pith/6J3DFGYFRAFTOLFHIH7YFU7VCM/action/replication_record"}},"created_at":"2026-07-05T10:48:36.832928+00:00","updated_at":"2026-07-05T10:48:36.832928+00:00"}