{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:CG2QVDPKJ4T3LFO4E3NSS7H5C2","short_pith_number":"pith:CG2QVDPK","schema_version":"1.0","canonical_sha256":"11b50a8dea4f27b595dc26db297cfd16be9bc8e9ab7e02ac304e1e9126f11b75","source":{"kind":"arxiv","id":"2312.03813","version":1},"attestation_state":"computed","paper":{"title":"Improving Activation Steering in Language Models with Mean-Centring","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Dylan Cope, Murray Shanahan, Nandi Schoots, Ole Jorgensen","submitted_at":"2023-12-06T18:27:07Z","abstract_excerpt":"Recent work in activation steering has demonstrated the potential to better control the outputs of Large Language Models (LLMs), but it involves finding steering vectors. This is difficult because engineers do not typically know how features are represented in these models. We seek to address this issue by applying the idea of mean-centring to steering vectors. We find that taking the average of activations associated with a target dataset, and then subtracting the mean of all training activations, results in effective steering vectors. We test this method on a variety of models on natural lan"},"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":"2312.03813","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-12-06T18:27:07Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"5a042eb3b5019172451afb6ec28db73c1ca736b3365b900683ff8125955d4dc1","abstract_canon_sha256":"74a98faadace7ecd58cc44d03c6d9d5524a2bcdac297f7172d507c7fd39add1e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:21:15.380658Z","signature_b64":"+6ItyK4Vb+uBKg69gEKyaKcJ/CEwVq8Ir1BdoJDYNtngXEpAZhWSdReiyDkKez6+l6Vid9RatKY9glvUsTGuCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"11b50a8dea4f27b595dc26db297cfd16be9bc8e9ab7e02ac304e1e9126f11b75","last_reissued_at":"2026-07-05T07:21:15.380158Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:21:15.380158Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Improving Activation Steering in Language Models with Mean-Centring","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Dylan Cope, Murray Shanahan, Nandi Schoots, Ole Jorgensen","submitted_at":"2023-12-06T18:27:07Z","abstract_excerpt":"Recent work in activation steering has demonstrated the potential to better control the outputs of Large Language Models (LLMs), but it involves finding steering vectors. This is difficult because engineers do not typically know how features are represented in these models. We seek to address this issue by applying the idea of mean-centring to steering vectors. We find that taking the average of activations associated with a target dataset, and then subtracting the mean of all training activations, results in effective steering vectors. We test this method on a variety of models on natural lan"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2312.03813","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/2312.03813/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":"2312.03813","created_at":"2026-07-05T07:21:15.380222+00:00"},{"alias_kind":"arxiv_version","alias_value":"2312.03813v1","created_at":"2026-07-05T07:21:15.380222+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2312.03813","created_at":"2026-07-05T07:21:15.380222+00:00"},{"alias_kind":"pith_short_12","alias_value":"CG2QVDPKJ4T3","created_at":"2026-07-05T07:21:15.380222+00:00"},{"alias_kind":"pith_short_16","alias_value":"CG2QVDPKJ4T3LFO4","created_at":"2026-07-05T07:21:15.380222+00:00"},{"alias_kind":"pith_short_8","alias_value":"CG2QVDPK","created_at":"2026-07-05T07:21:15.380222+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.11599","citing_title":"When is Your LLM Steerable?","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01844","citing_title":"The Cylindrical Representation Hypothesis for Language Model Steering","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2406.11717","citing_title":"Refusal in Language Models Is Mediated by a Single Direction","ref_index":142,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06196","citing_title":"The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models","ref_index":61,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05715","citing_title":"Decodable but Not Corrected by Fixed Residual-Stream Linear Steering: Evidence from Medical LLM Failure Regimes","ref_index":34,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CG2QVDPKJ4T3LFO4E3NSS7H5C2","json":"https://pith.science/pith/CG2QVDPKJ4T3LFO4E3NSS7H5C2.json","graph_json":"https://pith.science/api/pith-number/CG2QVDPKJ4T3LFO4E3NSS7H5C2/graph.json","events_json":"https://pith.science/api/pith-number/CG2QVDPKJ4T3LFO4E3NSS7H5C2/events.json","paper":"https://pith.science/paper/CG2QVDPK"},"agent_actions":{"view_html":"https://pith.science/pith/CG2QVDPKJ4T3LFO4E3NSS7H5C2","download_json":"https://pith.science/pith/CG2QVDPKJ4T3LFO4E3NSS7H5C2.json","view_paper":"https://pith.science/paper/CG2QVDPK","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2312.03813&json=true","fetch_graph":"https://pith.science/api/pith-number/CG2QVDPKJ4T3LFO4E3NSS7H5C2/graph.json","fetch_events":"https://pith.science/api/pith-number/CG2QVDPKJ4T3LFO4E3NSS7H5C2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CG2QVDPKJ4T3LFO4E3NSS7H5C2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CG2QVDPKJ4T3LFO4E3NSS7H5C2/action/storage_attestation","attest_author":"https://pith.science/pith/CG2QVDPKJ4T3LFO4E3NSS7H5C2/action/author_attestation","sign_citation":"https://pith.science/pith/CG2QVDPKJ4T3LFO4E3NSS7H5C2/action/citation_signature","submit_replication":"https://pith.science/pith/CG2QVDPKJ4T3LFO4E3NSS7H5C2/action/replication_record"}},"created_at":"2026-07-05T07:21:15.380222+00:00","updated_at":"2026-07-05T07:21:15.380222+00:00"}