{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:YKH7BFTMRYGQRT2GW36XLQKOS5","short_pith_number":"pith:YKH7BFTM","schema_version":"1.0","canonical_sha256":"c28ff0966c8e0d08cf46b6fd75c14e976a9d2f44ac0244519608f327f0f796bd","source":{"kind":"arxiv","id":"2502.09674","version":4},"attestation_state":"computed","paper":{"title":"The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Haining Yu, Qiguang Chen, Wenbo Pan, Xiangyang Zhou, Xiaohua Jia, Zhichao Liu","submitted_at":"2025-02-13T06:39:22Z","abstract_excerpt":"Large Language Models' safety-aligned behaviors, such as refusing harmful queries, can be represented by linear directions in activation space. Previous research modeled safety behavior with a single direction, limiting mechanistic understanding to an isolated safety feature. In this work, we discover that safety-aligned behavior is jointly controlled by multi-dimensional directions. Namely, we study the vector space of representation shifts during safety fine-tuning on Llama 3 8B for refusing jailbreaks. By studying orthogonal directions in the space, we first find that a dominant direction 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":"2502.09674","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-02-13T06:39:22Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"13ee7bd8a53f281f32b4eed8928919a7829518e08146142629a4abc9626e7c3e","abstract_canon_sha256":"4a581e9522c89912cb6b850e11ecb4ac20b21c7e097f2126d1e909d294ec9a86"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:10:04.895690Z","signature_b64":"oMS/52DTuWk+ZYjAfU5P7GuoQSr2lxVX0Zh0r6ixf5Tu7h+skIxSTP6i0uhOUsi9PEEOfWVI5kt6dC+PYvqXCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c28ff0966c8e0d08cf46b6fd75c14e976a9d2f44ac0244519608f327f0f796bd","last_reissued_at":"2026-07-05T11:10:04.895155Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:10:04.895155Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Haining Yu, Qiguang Chen, Wenbo Pan, Xiangyang Zhou, Xiaohua Jia, Zhichao Liu","submitted_at":"2025-02-13T06:39:22Z","abstract_excerpt":"Large Language Models' safety-aligned behaviors, such as refusing harmful queries, can be represented by linear directions in activation space. Previous research modeled safety behavior with a single direction, limiting mechanistic understanding to an isolated safety feature. In this work, we discover that safety-aligned behavior is jointly controlled by multi-dimensional directions. Namely, we study the vector space of representation shifts during safety fine-tuning on Llama 3 8B for refusing jailbreaks. By studying orthogonal directions in the space, we first find that a dominant direction g"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.09674","kind":"arxiv","version":4},"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/2502.09674/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":"2502.09674","created_at":"2026-07-05T11:10:04.895214+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.09674v4","created_at":"2026-07-05T11:10:04.895214+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.09674","created_at":"2026-07-05T11:10:04.895214+00:00"},{"alias_kind":"pith_short_12","alias_value":"YKH7BFTMRYGQ","created_at":"2026-07-05T11:10:04.895214+00:00"},{"alias_kind":"pith_short_16","alias_value":"YKH7BFTMRYGQRT2G","created_at":"2026-07-05T11:10:04.895214+00:00"},{"alias_kind":"pith_short_8","alias_value":"YKH7BFTM","created_at":"2026-07-05T11:10:04.895214+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":11,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2607.01854","citing_title":"Has This Checkpoint Been Abliterated? A Two-Signal Audit and Its Failure Map","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02530","citing_title":"SafeSteer: Localized On-Policy Distillation for Efficient Safety Alignment","ref_index":50,"is_internal_anchor":false},{"citing_arxiv_id":"2606.01196","citing_title":"Low-Resource Safety Failures Are Action Failures, Not Representation Failures","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17173","citing_title":"Why Do Safety Guardrails Degrade Across Languages?","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2506.01770","citing_title":"ReGA: Model-Based Safeguard for LLMs via Representation-Guided Abstraction","ref_index":66,"is_internal_anchor":false},{"citing_arxiv_id":"2602.02280","citing_title":"RACC: Representation-Aware Coverage Criteria for LLM Safety Testing","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14514","citing_title":"Defenses at Odds: Measuring and Explaining Defense Conflicts in Large Language Models","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12726","citing_title":"Before the Last Token: Diagnosing Final-Token Safety Probe Failures","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11679","citing_title":"Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11679","citing_title":"Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17691","citing_title":"SafeAnchor: Preventing Cumulative Safety Erosion in Continual Domain Adaptation of Large Language Models","ref_index":20,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/YKH7BFTMRYGQRT2GW36XLQKOS5","json":"https://pith.science/pith/YKH7BFTMRYGQRT2GW36XLQKOS5.json","graph_json":"https://pith.science/api/pith-number/YKH7BFTMRYGQRT2GW36XLQKOS5/graph.json","events_json":"https://pith.science/api/pith-number/YKH7BFTMRYGQRT2GW36XLQKOS5/events.json","paper":"https://pith.science/paper/YKH7BFTM"},"agent_actions":{"view_html":"https://pith.science/pith/YKH7BFTMRYGQRT2GW36XLQKOS5","download_json":"https://pith.science/pith/YKH7BFTMRYGQRT2GW36XLQKOS5.json","view_paper":"https://pith.science/paper/YKH7BFTM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.09674&json=true","fetch_graph":"https://pith.science/api/pith-number/YKH7BFTMRYGQRT2GW36XLQKOS5/graph.json","fetch_events":"https://pith.science/api/pith-number/YKH7BFTMRYGQRT2GW36XLQKOS5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/YKH7BFTMRYGQRT2GW36XLQKOS5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/YKH7BFTMRYGQRT2GW36XLQKOS5/action/storage_attestation","attest_author":"https://pith.science/pith/YKH7BFTMRYGQRT2GW36XLQKOS5/action/author_attestation","sign_citation":"https://pith.science/pith/YKH7BFTMRYGQRT2GW36XLQKOS5/action/citation_signature","submit_replication":"https://pith.science/pith/YKH7BFTMRYGQRT2GW36XLQKOS5/action/replication_record"}},"created_at":"2026-07-05T11:10:04.895214+00:00","updated_at":"2026-07-05T11:10:04.895214+00:00"}