{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:4QQ672QDBKWRYQU35ZUYZ5MMRD","short_pith_number":"pith:4QQ672QD","schema_version":"1.0","canonical_sha256":"e421efea030aad1c429bee698cf58c88c7705e00a5e7347710bd1e22608ea81f","source":{"kind":"arxiv","id":"2411.08790","version":1},"attestation_state":"computed","paper":{"title":"Can sparse autoencoders be used to decompose and interpret steering vectors?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Adam Mahdi, Harry Mayne, Yushi Yang","submitted_at":"2024-11-13T17:16:48Z","abstract_excerpt":"Steering vectors are a promising approach to control the behaviour of large language models. However, their underlying mechanisms remain poorly understood. While sparse autoencoders (SAEs) may offer a potential method to interpret steering vectors, recent findings show that SAE-reconstructed vectors often lack the steering properties of the original vectors. This paper investigates why directly applying SAEs to steering vectors yields misleading decompositions, identifying two reasons: (1) steering vectors fall outside the input distribution for which SAEs are designed, and (2) steering vector"},"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":"2411.08790","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-11-13T17:16:48Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"47c659c61e1bd00efe5c4fb2ec304aae51e277dbcaa1ec1edd3a5c3fbc0f3a4c","abstract_canon_sha256":"c3ea27355a2ddb792c379bb6b3aa1d0c113e3e2a56f30c2f76f3e95bd290d7e5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:35:02.510435Z","signature_b64":"Eaw00VRDOBhDXlOJjRXhR7CZxBPoF9Sr7ggpoyhC2bOEkjdGpw8XaQ0D+P/O//0L1u06HtSMoAgHnJgmni08AA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e421efea030aad1c429bee698cf58c88c7705e00a5e7347710bd1e22608ea81f","last_reissued_at":"2026-07-05T09:35:02.509978Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:35:02.509978Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Can sparse autoencoders be used to decompose and interpret steering vectors?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Adam Mahdi, Harry Mayne, Yushi Yang","submitted_at":"2024-11-13T17:16:48Z","abstract_excerpt":"Steering vectors are a promising approach to control the behaviour of large language models. However, their underlying mechanisms remain poorly understood. While sparse autoencoders (SAEs) may offer a potential method to interpret steering vectors, recent findings show that SAE-reconstructed vectors often lack the steering properties of the original vectors. This paper investigates why directly applying SAEs to steering vectors yields misleading decompositions, identifying two reasons: (1) steering vectors fall outside the input distribution for which SAEs are designed, and (2) steering vector"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2411.08790","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/2411.08790/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":"2411.08790","created_at":"2026-07-05T09:35:02.510035+00:00"},{"alias_kind":"arxiv_version","alias_value":"2411.08790v1","created_at":"2026-07-05T09:35:02.510035+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2411.08790","created_at":"2026-07-05T09:35:02.510035+00:00"},{"alias_kind":"pith_short_12","alias_value":"4QQ672QDBKWR","created_at":"2026-07-05T09:35:02.510035+00:00"},{"alias_kind":"pith_short_16","alias_value":"4QQ672QDBKWRYQU3","created_at":"2026-07-05T09:35:02.510035+00:00"},{"alias_kind":"pith_short_8","alias_value":"4QQ672QD","created_at":"2026-07-05T09:35:02.510035+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2505.15524","citing_title":"Evaluate Bias without Manual Test Sets: A Concept Representation Perspective for LLMs","ref_index":14,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4QQ672QDBKWRYQU35ZUYZ5MMRD","json":"https://pith.science/pith/4QQ672QDBKWRYQU35ZUYZ5MMRD.json","graph_json":"https://pith.science/api/pith-number/4QQ672QDBKWRYQU35ZUYZ5MMRD/graph.json","events_json":"https://pith.science/api/pith-number/4QQ672QDBKWRYQU35ZUYZ5MMRD/events.json","paper":"https://pith.science/paper/4QQ672QD"},"agent_actions":{"view_html":"https://pith.science/pith/4QQ672QDBKWRYQU35ZUYZ5MMRD","download_json":"https://pith.science/pith/4QQ672QDBKWRYQU35ZUYZ5MMRD.json","view_paper":"https://pith.science/paper/4QQ672QD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2411.08790&json=true","fetch_graph":"https://pith.science/api/pith-number/4QQ672QDBKWRYQU35ZUYZ5MMRD/graph.json","fetch_events":"https://pith.science/api/pith-number/4QQ672QDBKWRYQU35ZUYZ5MMRD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4QQ672QDBKWRYQU35ZUYZ5MMRD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4QQ672QDBKWRYQU35ZUYZ5MMRD/action/storage_attestation","attest_author":"https://pith.science/pith/4QQ672QDBKWRYQU35ZUYZ5MMRD/action/author_attestation","sign_citation":"https://pith.science/pith/4QQ672QDBKWRYQU35ZUYZ5MMRD/action/citation_signature","submit_replication":"https://pith.science/pith/4QQ672QDBKWRYQU35ZUYZ5MMRD/action/replication_record"}},"created_at":"2026-07-05T09:35:02.510035+00:00","updated_at":"2026-07-05T09:35:02.510035+00:00"}