{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:K2YZJN2U6ZLNQGH6NGRWOLLK5U","short_pith_number":"pith:K2YZJN2U","schema_version":"1.0","canonical_sha256":"56b194b754f656d818fe69a3672d6aed0a210f8de391e093330c61f7b9bc1739","source":{"kind":"arxiv","id":"2603.17839","version":3},"attestation_state":"computed","paper":{"title":"How do LLMs Compute Verbal Confidence","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Arthur Conmy, Dharshan Kumaran, Federico Barbero, Petar Veli\\v{c}kovi\\'c, Simon Osindero, Viorica Patraucean","submitted_at":"2026-03-18T15:31:43Z","abstract_excerpt":"Verbal confidence -- prompting LLMs to state their confidence as a number or category -- is widely used to extract uncertainty estimates from black-box models. However, how LLMs internally generate such scores remains unknown. We address two questions: first, when confidence is computed -- just-in-time when requested, or automatically during answer generation and cached for later retrieval; and second, what verbal confidence represents -- token log-probabilities, or a richer evaluation of answer quality? Focusing on Gemma 3 27B (across TriviaQA, BigMath, and MMLU), Qwen 2.5 7B, and the reasoni"},"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":"2603.17839","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-03-18T15:31:43Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"0b4ff1b7d12e9cc61f35d8fcb6f981bc9f6f6a3d78372f8579acbc9abecf4e3f","abstract_canon_sha256":"92e2b51876eb8b141161dfb9d091eb21a04c9b10eb323bf6ccc804b60e52144e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T01:05:11.401669Z","signature_b64":"ikJxZIfN2+W3DkYwhnyFLQhKAlOw36lnfqucsI2nrwjzvDEHg3mVSO/pjnM/Rz2+mDUMxU+MaT0UF/3KBEJiDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"56b194b754f656d818fe69a3672d6aed0a210f8de391e093330c61f7b9bc1739","last_reissued_at":"2026-05-20T01:05:11.400901Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T01:05:11.400901Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"How do LLMs Compute Verbal Confidence","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Arthur Conmy, Dharshan Kumaran, Federico Barbero, Petar Veli\\v{c}kovi\\'c, Simon Osindero, Viorica Patraucean","submitted_at":"2026-03-18T15:31:43Z","abstract_excerpt":"Verbal confidence -- prompting LLMs to state their confidence as a number or category -- is widely used to extract uncertainty estimates from black-box models. However, how LLMs internally generate such scores remains unknown. We address two questions: first, when confidence is computed -- just-in-time when requested, or automatically during answer generation and cached for later retrieval; and second, what verbal confidence represents -- token log-probabilities, or a richer evaluation of answer quality? Focusing on Gemma 3 27B (across TriviaQA, BigMath, and MMLU), Qwen 2.5 7B, and the reasoni"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.17839","kind":"arxiv","version":3},"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/2603.17839/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":"2603.17839","created_at":"2026-05-20T01:05:11.401033+00:00"},{"alias_kind":"arxiv_version","alias_value":"2603.17839v3","created_at":"2026-05-20T01:05:11.401033+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.17839","created_at":"2026-05-20T01:05:11.401033+00:00"},{"alias_kind":"pith_short_12","alias_value":"K2YZJN2U6ZLN","created_at":"2026-05-20T01:05:11.401033+00:00"},{"alias_kind":"pith_short_16","alias_value":"K2YZJN2U6ZLNQGH6","created_at":"2026-05-20T01:05:11.401033+00:00"},{"alias_kind":"pith_short_8","alias_value":"K2YZJN2U","created_at":"2026-05-20T01:05:11.401033+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":6,"sample":[{"citing_arxiv_id":"2605.20745","citing_title":"The Hidden Signal of Verifier Strictness: Controlling and Improving Step-Wise Verification via Selective Latent Steering","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2605.11235","citing_title":"Internalizing Curriculum Judgment for LLM Reinforcement Fine-Tuning","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2604.24070","citing_title":"Distilling Self-Consistency into Verbal Confidence: A Pre-Registered Negative Result and Post-Hoc Rescue on Gemma 3 4B","ref_index":9,"is_internal_anchor":true},{"citing_arxiv_id":"2604.22215","citing_title":"Verbal Confidence Saturation in 3-9B Open-Weight Instruction-Tuned LLMs: A Pre-Registered Psychometric Validity Screen","ref_index":12,"is_internal_anchor":true},{"citing_arxiv_id":"2604.22271","citing_title":"How LLMs Detect and Correct Their Own Errors: The Role of Internal Confidence Signals","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05851","citing_title":"Hypothesis generation and updating in large language models","ref_index":25,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/K2YZJN2U6ZLNQGH6NGRWOLLK5U","json":"https://pith.science/pith/K2YZJN2U6ZLNQGH6NGRWOLLK5U.json","graph_json":"https://pith.science/api/pith-number/K2YZJN2U6ZLNQGH6NGRWOLLK5U/graph.json","events_json":"https://pith.science/api/pith-number/K2YZJN2U6ZLNQGH6NGRWOLLK5U/events.json","paper":"https://pith.science/paper/K2YZJN2U"},"agent_actions":{"view_html":"https://pith.science/pith/K2YZJN2U6ZLNQGH6NGRWOLLK5U","download_json":"https://pith.science/pith/K2YZJN2U6ZLNQGH6NGRWOLLK5U.json","view_paper":"https://pith.science/paper/K2YZJN2U","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2603.17839&json=true","fetch_graph":"https://pith.science/api/pith-number/K2YZJN2U6ZLNQGH6NGRWOLLK5U/graph.json","fetch_events":"https://pith.science/api/pith-number/K2YZJN2U6ZLNQGH6NGRWOLLK5U/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/K2YZJN2U6ZLNQGH6NGRWOLLK5U/action/timestamp_anchor","attest_storage":"https://pith.science/pith/K2YZJN2U6ZLNQGH6NGRWOLLK5U/action/storage_attestation","attest_author":"https://pith.science/pith/K2YZJN2U6ZLNQGH6NGRWOLLK5U/action/author_attestation","sign_citation":"https://pith.science/pith/K2YZJN2U6ZLNQGH6NGRWOLLK5U/action/citation_signature","submit_replication":"https://pith.science/pith/K2YZJN2U6ZLNQGH6NGRWOLLK5U/action/replication_record"}},"created_at":"2026-05-20T01:05:11.401033+00:00","updated_at":"2026-05-20T01:05:11.401033+00:00"}