{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:44MJ5GBCUQWUO5A4V2HHQUI6WA","short_pith_number":"pith:44MJ5GBC","schema_version":"1.0","canonical_sha256":"e7189e9822a42d47741cae8e78511eb01fbf000447c9c9c3f60d28efe902de94","source":{"kind":"arxiv","id":"2605.01844","version":2},"attestation_state":"computed","paper":{"title":"The Cylindrical Representation Hypothesis for Language Model Steering","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"The Cylindrical Representation Hypothesis models concept representations in LLMs as a central axis for concept presence surrounded by a normal plane containing sensitive sectors that control activation ease, explaining steering unpredictability.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Akash Ghosh, Chenxi Wang, Fengxian Ji, Jinghui Zhang, Lang Gao, Preslav Nakov, Wei Liu, Xiuying Chen, Youssef Mohamed, Zirui Song","submitted_at":"2026-05-03T12:26:13Z","abstract_excerpt":"Steering is a widely used technique for controlling large language models, yet its effects are often unstable and hard to predict. Existing theoretical accounts are largely based on the Linear Representation Hypothesis (LRH). While LRH assumes that concepts can be orthogonalized for lossless control, this idealized mapping fails in real representations and cannot account for the observed unpredictability of steering. By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogona"},"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":"2605.01844","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-03T12:26:13Z","cross_cats_sorted":[],"title_canon_sha256":"b89bd423b4f4bea0de39702db40cf8ee7eac111366eaf5e41628ece2dd9495c8","abstract_canon_sha256":"664ce3267c40e97b537c90a951368df9b7911e8843b899d9522243a9079d5399"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-05T01:15:25.057436Z","signature_b64":"jTBG71hk03U3YNqETmMZJ4fcu3ORqVKl0ibqamT8HjRYOwurW6D796g881yPtt4RoE352//uemQx6/8b1OkaAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e7189e9822a42d47741cae8e78511eb01fbf000447c9c9c3f60d28efe902de94","last_reissued_at":"2026-06-05T01:15:25.056925Z","signature_status":"signed_v1","first_computed_at":"2026-06-05T01:15:25.056925Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Cylindrical Representation Hypothesis for Language Model Steering","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"The Cylindrical Representation Hypothesis models concept representations in LLMs as a central axis for concept presence surrounded by a normal plane containing sensitive sectors that control activation ease, explaining steering unpredictability.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Akash Ghosh, Chenxi Wang, Fengxian Ji, Jinghui Zhang, Lang Gao, Preslav Nakov, Wei Liu, Xiuying Chen, Youssef Mohamed, Zirui Song","submitted_at":"2026-05-03T12:26:13Z","abstract_excerpt":"Steering is a widely used technique for controlling large language models, yet its effects are often unstable and hard to predict. Existing theoretical accounts are largely based on the Linear Representation Hypothesis (LRH). While LRH assumes that concepts can be orthogonalized for lossless control, this idealized mapping fails in real representations and cannot account for the observed unpredictability of steering. By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogona"},"claims":{"count":3,"items":[{"kind":"strongest_claim","text":"By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogonal structure. We formalize this as the Cylindrical Representation Hypothesis (CRH).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That overlapping concept contributions produce a reliably identifiable normal plane from difference vectors while the sensitive sector within that plane remains intrinsically unidentifiable, introducing unavoidable uncertainty at the sector level.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"The Cylindrical Representation Hypothesis models concept representations in LLMs as a central axis for concept presence surrounded by a normal plane containing sensitive sectors that control activation ease, explaining steering unpredictability.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"}],"snapshot_sha256":"80de4df37389cd9590c7efd1b35b6133e0a08005dc46df388e4426e1dcca40a8"},"source":{"id":"2605.01844","kind":"arxiv","version":2},"verdict":{"id":"fdcb3bb8-30fb-4f44-b3cc-26e97cefd5ea","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-09T13:43:16.319471Z","strongest_claim":"By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogonal structure. We formalize this as the Cylindrical Representation Hypothesis (CRH).","one_line_summary":"The Cylindrical Representation Hypothesis models concept representations in LLMs as a central axis for concept presence surrounded by a normal plane containing sensitive sectors that control activation ease, explaining steering unpredictability.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That overlapping concept contributions produce a reliably identifiable normal plane from difference vectors while the sensitive sector within that plane remains intrinsically unidentifiable, introducing unavoidable uncertainty at the sector level.","pith_extraction_headline":""},"integrity":{"clean":false,"summary":{"advisory":0,"critical":1,"by_detector":{"ai_meta_artifact":{"total":1,"advisory":0,"critical":1,"informational":0}},"informational":0},"endpoint":"/pith/2605.01844/integrity.json","findings":[{"note":"Verbatim AI-assistant artifact present in paper body: matched pattern 'certainly_here_is'. The matched span is the literal evidence.","detector":"ai_meta_artifact","severity":"critical","ref_index":null,"audited_at":"2026-05-20T17:35:01.483844Z","detected_doi":null,"finding_type":"certainly_here_is","verdict_class":"incontrovertible","detected_arxiv_id":null}],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T17:35:01.483844Z","status":"completed","version":"1.0.0","findings_count":1}],"snapshot_sha256":"a025282c910e9616d8bd323fc7f750f4ec4c346c601f795e21f86ad953b61d9d"},"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":"2605.01844","created_at":"2026-06-05T01:15:25.056995+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.01844v2","created_at":"2026-06-05T01:15:25.056995+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.01844","created_at":"2026-06-05T01:15:25.056995+00:00"},{"alias_kind":"pith_short_12","alias_value":"44MJ5GBCUQWU","created_at":"2026-06-05T01:15:25.056995+00:00"},{"alias_kind":"pith_short_16","alias_value":"44MJ5GBCUQWUO5A4","created_at":"2026-06-05T01:15:25.056995+00:00"},{"alias_kind":"pith_short_8","alias_value":"44MJ5GBC","created_at":"2026-06-05T01:15:25.056995+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"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":20,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/44MJ5GBCUQWUO5A4V2HHQUI6WA","json":"https://pith.science/pith/44MJ5GBCUQWUO5A4V2HHQUI6WA.json","graph_json":"https://pith.science/api/pith-number/44MJ5GBCUQWUO5A4V2HHQUI6WA/graph.json","events_json":"https://pith.science/api/pith-number/44MJ5GBCUQWUO5A4V2HHQUI6WA/events.json","paper":"https://pith.science/paper/44MJ5GBC"},"agent_actions":{"view_html":"https://pith.science/pith/44MJ5GBCUQWUO5A4V2HHQUI6WA","download_json":"https://pith.science/pith/44MJ5GBCUQWUO5A4V2HHQUI6WA.json","view_paper":"https://pith.science/paper/44MJ5GBC","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.01844&json=true","fetch_graph":"https://pith.science/api/pith-number/44MJ5GBCUQWUO5A4V2HHQUI6WA/graph.json","fetch_events":"https://pith.science/api/pith-number/44MJ5GBCUQWUO5A4V2HHQUI6WA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/44MJ5GBCUQWUO5A4V2HHQUI6WA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/44MJ5GBCUQWUO5A4V2HHQUI6WA/action/storage_attestation","attest_author":"https://pith.science/pith/44MJ5GBCUQWUO5A4V2HHQUI6WA/action/author_attestation","sign_citation":"https://pith.science/pith/44MJ5GBCUQWUO5A4V2HHQUI6WA/action/citation_signature","submit_replication":"https://pith.science/pith/44MJ5GBCUQWUO5A4V2HHQUI6WA/action/replication_record"}},"created_at":"2026-06-05T01:15:25.056995+00:00","updated_at":"2026-06-05T01:15:25.056995+00:00"}