{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:STOKPVAD7EPF5LCG4CEXFEECQP","short_pith_number":"pith:STOKPVAD","schema_version":"1.0","canonical_sha256":"94dca7d403f91e5eac46e08972908283dc73010aff5e7c5182cc4c6351baf8df","source":{"kind":"arxiv","id":"2505.22637","version":1},"attestation_state":"computed","paper":{"title":"Understanding (Un)Reliability of Steering Vectors in Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Carsten Eickhoff, David Krueger, Dmitrii Krasheninnikov, Joschka Braun, Seyed Ali Bahrainian","submitted_at":"2025-05-28T17:53:31Z","abstract_excerpt":"Steering vectors are a lightweight method to control language model behavior by adding a learned bias to the activations at inference time. Although steering demonstrates promising performance, recent work shows that it can be unreliable or even counterproductive in some cases. This paper studies the influence of prompt types and the geometry of activation differences on steering reliability. First, we find that all seven prompt types used in our experiments produce a net positive steering effect, but exhibit high variance across samples, and often give an effect opposite of the desired one. N"},"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":"2505.22637","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-05-28T17:53:31Z","cross_cats_sorted":[],"title_canon_sha256":"f9c4670352a5b4ade7de80ff884fa804eb08622df06f11933389fb1ccca42837","abstract_canon_sha256":"93c11547f668287b36be561896abd19feaf06a51e2d25a4f138edb9b7077f7da"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:11:28.204342Z","signature_b64":"ehqN0z0mOZrPG1JaYMT9wLSSED0dbCHEddNgx1vS5OqbWygwl5oqYmH0u6aQZ4ce0aSjOu8/WbsrLGoCWpQ/BQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"94dca7d403f91e5eac46e08972908283dc73010aff5e7c5182cc4c6351baf8df","last_reissued_at":"2026-07-05T11:11:28.203783Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:11:28.203783Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Understanding (Un)Reliability of Steering Vectors in Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Carsten Eickhoff, David Krueger, Dmitrii Krasheninnikov, Joschka Braun, Seyed Ali Bahrainian","submitted_at":"2025-05-28T17:53:31Z","abstract_excerpt":"Steering vectors are a lightweight method to control language model behavior by adding a learned bias to the activations at inference time. Although steering demonstrates promising performance, recent work shows that it can be unreliable or even counterproductive in some cases. This paper studies the influence of prompt types and the geometry of activation differences on steering reliability. First, we find that all seven prompt types used in our experiments produce a net positive steering effect, but exhibit high variance across samples, and often give an effect opposite of the desired one. N"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.22637","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/2505.22637/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":"2505.22637","created_at":"2026-07-05T11:11:28.203844+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.22637v1","created_at":"2026-07-05T11:11:28.203844+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.22637","created_at":"2026-07-05T11:11:28.203844+00:00"},{"alias_kind":"pith_short_12","alias_value":"STOKPVAD7EPF","created_at":"2026-07-05T11:11:28.203844+00:00"},{"alias_kind":"pith_short_16","alias_value":"STOKPVAD7EPF5LCG","created_at":"2026-07-05T11:11:28.203844+00:00"},{"alias_kind":"pith_short_8","alias_value":"STOKPVAD","created_at":"2026-07-05T11:11:28.203844+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":8,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26155","citing_title":"Detecting and Controlling Sycophancy with Cascading Linear Features","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07696","citing_title":"Adversarial Robustness of Activation Steering in Large Language Models","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05194","citing_title":"Temporal Preference Concepts and their Functions in a Large Language Model","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2605.16362","citing_title":"When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2605.16362","citing_title":"When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10664","citing_title":"Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10664","citing_title":"Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02946","citing_title":"RouteHijack: Routing-Aware Attack on Mixture-of-Experts LLMs","ref_index":5,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/STOKPVAD7EPF5LCG4CEXFEECQP","json":"https://pith.science/pith/STOKPVAD7EPF5LCG4CEXFEECQP.json","graph_json":"https://pith.science/api/pith-number/STOKPVAD7EPF5LCG4CEXFEECQP/graph.json","events_json":"https://pith.science/api/pith-number/STOKPVAD7EPF5LCG4CEXFEECQP/events.json","paper":"https://pith.science/paper/STOKPVAD"},"agent_actions":{"view_html":"https://pith.science/pith/STOKPVAD7EPF5LCG4CEXFEECQP","download_json":"https://pith.science/pith/STOKPVAD7EPF5LCG4CEXFEECQP.json","view_paper":"https://pith.science/paper/STOKPVAD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.22637&json=true","fetch_graph":"https://pith.science/api/pith-number/STOKPVAD7EPF5LCG4CEXFEECQP/graph.json","fetch_events":"https://pith.science/api/pith-number/STOKPVAD7EPF5LCG4CEXFEECQP/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/STOKPVAD7EPF5LCG4CEXFEECQP/action/timestamp_anchor","attest_storage":"https://pith.science/pith/STOKPVAD7EPF5LCG4CEXFEECQP/action/storage_attestation","attest_author":"https://pith.science/pith/STOKPVAD7EPF5LCG4CEXFEECQP/action/author_attestation","sign_citation":"https://pith.science/pith/STOKPVAD7EPF5LCG4CEXFEECQP/action/citation_signature","submit_replication":"https://pith.science/pith/STOKPVAD7EPF5LCG4CEXFEECQP/action/replication_record"}},"created_at":"2026-07-05T11:11:28.203844+00:00","updated_at":"2026-07-05T11:11:28.203844+00:00"}