{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:4F72O4EYLY3QR53TN3WLASQBPT","short_pith_number":"pith:4F72O4EY","schema_version":"1.0","canonical_sha256":"e17fa770985e3708f7736eecb04a017ce7bffa2c06c12568dbc8ca4eee4a9b00","source":{"kind":"arxiv","id":"2505.21399","version":1},"attestation_state":"computed","paper":{"title":"Factual Self-Awareness in Language Models: Representation, Robustness, and Scaling","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Hovhannes Tamoyan, Iryna Gurevych, Subhabrata Dutta","submitted_at":"2025-05-27T16:24:02Z","abstract_excerpt":"Factual incorrectness in generated content is one of the primary concerns in ubiquitous deployment of large language models (LLMs). Prior findings suggest LLMs can (sometimes) detect factual incorrectness in their generated content (i.e., fact-checking post-generation). In this work, we provide evidence supporting the presence of LLMs' internal compass that dictate the correctness of factual recall at the time of generation. We demonstrate that for a given subject entity and a relation, LLMs internally encode linear features in the Transformer's residual stream that dictate whether it will be "},"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":"2505.21399","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-05-27T16:24:02Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"9469b7c63f8a0401a59db7944c456363e8f2116635ba7f6864fbf931952ed23f","abstract_canon_sha256":"f45e0515fe8e4542a9171cb2694e53e63adfc5050bac69c0afdd84b3bc77d6d0"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:10:41.428345Z","signature_b64":"jPDQiIAxloZBPJjr+viTDKz2X4SxqJYtiWRt2QQQHmpRInnNZNbO39SusVTpG4dHiniuhMug04eFecqvf41vAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e17fa770985e3708f7736eecb04a017ce7bffa2c06c12568dbc8ca4eee4a9b00","last_reissued_at":"2026-07-05T11:10:41.427750Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:10:41.427750Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Factual Self-Awareness in Language Models: Representation, Robustness, and Scaling","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Hovhannes Tamoyan, Iryna Gurevych, Subhabrata Dutta","submitted_at":"2025-05-27T16:24:02Z","abstract_excerpt":"Factual incorrectness in generated content is one of the primary concerns in ubiquitous deployment of large language models (LLMs). Prior findings suggest LLMs can (sometimes) detect factual incorrectness in their generated content (i.e., fact-checking post-generation). In this work, we provide evidence supporting the presence of LLMs' internal compass that dictate the correctness of factual recall at the time of generation. We demonstrate that for a given subject entity and a relation, LLMs internally encode linear features in the Transformer's residual stream that dictate whether it will be "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.21399","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/2505.21399/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":"2505.21399","created_at":"2026-07-05T11:10:41.427825+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.21399v1","created_at":"2026-07-05T11:10:41.427825+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.21399","created_at":"2026-07-05T11:10:41.427825+00:00"},{"alias_kind":"pith_short_12","alias_value":"4F72O4EYLY3Q","created_at":"2026-07-05T11:10:41.427825+00:00"},{"alias_kind":"pith_short_16","alias_value":"4F72O4EYLY3QR53T","created_at":"2026-07-05T11:10:41.427825+00:00"},{"alias_kind":"pith_short_8","alias_value":"4F72O4EY","created_at":"2026-07-05T11:10:41.427825+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2509.26238","citing_title":"Beyond Linear Probes: Dynamic Safety Monitoring for Language Models","ref_index":55,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4F72O4EYLY3QR53TN3WLASQBPT","json":"https://pith.science/pith/4F72O4EYLY3QR53TN3WLASQBPT.json","graph_json":"https://pith.science/api/pith-number/4F72O4EYLY3QR53TN3WLASQBPT/graph.json","events_json":"https://pith.science/api/pith-number/4F72O4EYLY3QR53TN3WLASQBPT/events.json","paper":"https://pith.science/paper/4F72O4EY"},"agent_actions":{"view_html":"https://pith.science/pith/4F72O4EYLY3QR53TN3WLASQBPT","download_json":"https://pith.science/pith/4F72O4EYLY3QR53TN3WLASQBPT.json","view_paper":"https://pith.science/paper/4F72O4EY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.21399&json=true","fetch_graph":"https://pith.science/api/pith-number/4F72O4EYLY3QR53TN3WLASQBPT/graph.json","fetch_events":"https://pith.science/api/pith-number/4F72O4EYLY3QR53TN3WLASQBPT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4F72O4EYLY3QR53TN3WLASQBPT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4F72O4EYLY3QR53TN3WLASQBPT/action/storage_attestation","attest_author":"https://pith.science/pith/4F72O4EYLY3QR53TN3WLASQBPT/action/author_attestation","sign_citation":"https://pith.science/pith/4F72O4EYLY3QR53TN3WLASQBPT/action/citation_signature","submit_replication":"https://pith.science/pith/4F72O4EYLY3QR53TN3WLASQBPT/action/replication_record"}},"created_at":"2026-07-05T11:10:41.427825+00:00","updated_at":"2026-07-05T11:10:41.427825+00:00"}