{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:PQKYF5EROXD6TIUCZF73S5OBGL","short_pith_number":"pith:PQKYF5ER","schema_version":"1.0","canonical_sha256":"7c1582f49175c7e9a282c97fb975c132fb0168fed8fe2fcab10d74b94d38a1f2","source":{"kind":"arxiv","id":"2312.10029","version":2},"attestation_state":"computed","paper":{"title":"Challenges with unsupervised LLM knowledge discovery","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Johannes Gasteiger, Rohin Shah, Sebastian Farquhar, Vikrant Varma, Vladimir Mikulik, Zachary Kenton","submitted_at":"2023-12-15T18:49:43Z","abstract_excerpt":"We show that existing unsupervised methods on large language model (LLM) activations do not discover knowledge -- instead they seem to discover whatever feature of the activations is most prominent. The idea behind unsupervised knowledge elicitation is that knowledge satisfies a consistency structure, which can be used to discover knowledge. We first prove theoretically that arbitrary features (not just knowledge) satisfy the consistency structure of a particular leading unsupervised knowledge-elicitation method, contrast-consistent search (Burns et al. - arXiv:2212.03827). We then present a s"},"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":"2312.10029","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2023-12-15T18:49:43Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"35f81195bdf14cbe5bb53c572389843e339b06d5cf5f84f44081eed871e36edf","abstract_canon_sha256":"eb8721bd48a1c4ad4b948617da5c0d9a4ea11fee35ce0156ec52b8c6f890f7b4"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:25:27.202050Z","signature_b64":"4x67VydWjOCUx9hnIZMXlWUeg+VapjGwB+AzykDqd43xIf4ojZw1pAEKrHKpmiP11vm0CCU4capAKxSrfz0vCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7c1582f49175c7e9a282c97fb975c132fb0168fed8fe2fcab10d74b94d38a1f2","last_reissued_at":"2026-07-05T07:25:27.201580Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:25:27.201580Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Challenges with unsupervised LLM knowledge discovery","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Johannes Gasteiger, Rohin Shah, Sebastian Farquhar, Vikrant Varma, Vladimir Mikulik, Zachary Kenton","submitted_at":"2023-12-15T18:49:43Z","abstract_excerpt":"We show that existing unsupervised methods on large language model (LLM) activations do not discover knowledge -- instead they seem to discover whatever feature of the activations is most prominent. The idea behind unsupervised knowledge elicitation is that knowledge satisfies a consistency structure, which can be used to discover knowledge. We first prove theoretically that arbitrary features (not just knowledge) satisfy the consistency structure of a particular leading unsupervised knowledge-elicitation method, contrast-consistent search (Burns et al. - arXiv:2212.03827). We then present a s"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2312.10029","kind":"arxiv","version":2},"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/2312.10029/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":"2312.10029","created_at":"2026-07-05T07:25:27.201643+00:00"},{"alias_kind":"arxiv_version","alias_value":"2312.10029v2","created_at":"2026-07-05T07:25:27.201643+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2312.10029","created_at":"2026-07-05T07:25:27.201643+00:00"},{"alias_kind":"pith_short_12","alias_value":"PQKYF5EROXD6","created_at":"2026-07-05T07:25:27.201643+00:00"},{"alias_kind":"pith_short_16","alias_value":"PQKYF5EROXD6TIUC","created_at":"2026-07-05T07:25:27.201643+00:00"},{"alias_kind":"pith_short_8","alias_value":"PQKYF5ER","created_at":"2026-07-05T07:25:27.201643+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.17229","citing_title":"Rift: A Conflict Signature for Deception in Language Models","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2606.27981","citing_title":"ToxiREX: A Dataset on Toxic REasoning in ConteXt","ref_index":137,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09563","citing_title":"PRISM: Recovering Instruction Sets from Language Model Activations","ref_index":15,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PQKYF5EROXD6TIUCZF73S5OBGL","json":"https://pith.science/pith/PQKYF5EROXD6TIUCZF73S5OBGL.json","graph_json":"https://pith.science/api/pith-number/PQKYF5EROXD6TIUCZF73S5OBGL/graph.json","events_json":"https://pith.science/api/pith-number/PQKYF5EROXD6TIUCZF73S5OBGL/events.json","paper":"https://pith.science/paper/PQKYF5ER"},"agent_actions":{"view_html":"https://pith.science/pith/PQKYF5EROXD6TIUCZF73S5OBGL","download_json":"https://pith.science/pith/PQKYF5EROXD6TIUCZF73S5OBGL.json","view_paper":"https://pith.science/paper/PQKYF5ER","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2312.10029&json=true","fetch_graph":"https://pith.science/api/pith-number/PQKYF5EROXD6TIUCZF73S5OBGL/graph.json","fetch_events":"https://pith.science/api/pith-number/PQKYF5EROXD6TIUCZF73S5OBGL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PQKYF5EROXD6TIUCZF73S5OBGL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PQKYF5EROXD6TIUCZF73S5OBGL/action/storage_attestation","attest_author":"https://pith.science/pith/PQKYF5EROXD6TIUCZF73S5OBGL/action/author_attestation","sign_citation":"https://pith.science/pith/PQKYF5EROXD6TIUCZF73S5OBGL/action/citation_signature","submit_replication":"https://pith.science/pith/PQKYF5EROXD6TIUCZF73S5OBGL/action/replication_record"}},"created_at":"2026-07-05T07:25:27.201643+00:00","updated_at":"2026-07-05T07:25:27.201643+00:00"}