{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:J2K3B3F3SQNZSZFVAOYIMOJPJA","short_pith_number":"pith:J2K3B3F3","schema_version":"1.0","canonical_sha256":"4e95b0ecbb941b9964b503b086392f48276027e448df46ffeba3e7ce3e242e05","source":{"kind":"arxiv","id":"2504.14379","version":2},"attestation_state":"computed","paper":{"title":"The Geometry of Self-Verification in a Task-Specific Reasoning Model","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Andrew Lee, Chris Wendler, Fernanda Vi\\'egas, Lihao Sun, Martin Wattenberg","submitted_at":"2025-04-19T18:40:51Z","abstract_excerpt":"How do reasoning models verify their own answers? We study this question by training a model using DeepSeek R1's recipe on the CountDown task. We leverage the fact that preference tuning leads to mode collapse, yielding a model that always produces highly structured chain-of-thought sequences. With this setup, we do top-down and bottom-up analyses to reverse-engineer how the model verifies its outputs. Top-down, we find Gated Linear Unit (GLU) weights encoding verification-related tokens, such as ``success'' or ``incorrect''. Bottom-up, we find that ``previous-token heads'' are mainly responsi"},"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":"2504.14379","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-04-19T18:40:51Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"259cc95fc978abbdd3662a1cb4970ffc0263b69bf1ef2e0533ccbe93296a0b31","abstract_canon_sha256":"9e8ab8154d64ce00033f8d706e7890b9ee54f12ffc2ec05dde6aa524e46dc405"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:01:35.207859Z","signature_b64":"2OGvHgnZu3WLcGCA9mfvJEOyjnj8PzGhxnxJHVwfe5u8SmSv+4VxVp5eE4KSesfTS/lcUx3zAEwhH+QR+HB4Cw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4e95b0ecbb941b9964b503b086392f48276027e448df46ffeba3e7ce3e242e05","last_reissued_at":"2026-07-05T11:01:35.206445Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:01:35.206445Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Geometry of Self-Verification in a Task-Specific Reasoning Model","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Andrew Lee, Chris Wendler, Fernanda Vi\\'egas, Lihao Sun, Martin Wattenberg","submitted_at":"2025-04-19T18:40:51Z","abstract_excerpt":"How do reasoning models verify their own answers? We study this question by training a model using DeepSeek R1's recipe on the CountDown task. We leverage the fact that preference tuning leads to mode collapse, yielding a model that always produces highly structured chain-of-thought sequences. With this setup, we do top-down and bottom-up analyses to reverse-engineer how the model verifies its outputs. Top-down, we find Gated Linear Unit (GLU) weights encoding verification-related tokens, such as ``success'' or ``incorrect''. Bottom-up, we find that ``previous-token heads'' are mainly responsi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.14379","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/2504.14379/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":"2504.14379","created_at":"2026-07-05T11:01:35.206553+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.14379v2","created_at":"2026-07-05T11:01:35.206553+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.14379","created_at":"2026-07-05T11:01:35.206553+00:00"},{"alias_kind":"pith_short_12","alias_value":"J2K3B3F3SQNZ","created_at":"2026-07-05T11:01:35.206553+00:00"},{"alias_kind":"pith_short_16","alias_value":"J2K3B3F3SQNZSZFV","created_at":"2026-07-05T11:01:35.206553+00:00"},{"alias_kind":"pith_short_8","alias_value":"J2K3B3F3","created_at":"2026-07-05T11:01:35.206553+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18284","citing_title":"Breaking the Solver Bottleneck: Training Task Generators at the Learnable Frontier","ref_index":108,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09967","citing_title":"Tensor Product Representation Probes Reveal Shared Structure Across Linear Directions","ref_index":12,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/J2K3B3F3SQNZSZFVAOYIMOJPJA","json":"https://pith.science/pith/J2K3B3F3SQNZSZFVAOYIMOJPJA.json","graph_json":"https://pith.science/api/pith-number/J2K3B3F3SQNZSZFVAOYIMOJPJA/graph.json","events_json":"https://pith.science/api/pith-number/J2K3B3F3SQNZSZFVAOYIMOJPJA/events.json","paper":"https://pith.science/paper/J2K3B3F3"},"agent_actions":{"view_html":"https://pith.science/pith/J2K3B3F3SQNZSZFVAOYIMOJPJA","download_json":"https://pith.science/pith/J2K3B3F3SQNZSZFVAOYIMOJPJA.json","view_paper":"https://pith.science/paper/J2K3B3F3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.14379&json=true","fetch_graph":"https://pith.science/api/pith-number/J2K3B3F3SQNZSZFVAOYIMOJPJA/graph.json","fetch_events":"https://pith.science/api/pith-number/J2K3B3F3SQNZSZFVAOYIMOJPJA/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/J2K3B3F3SQNZSZFVAOYIMOJPJA/action/timestamp_anchor","attest_storage":"https://pith.science/pith/J2K3B3F3SQNZSZFVAOYIMOJPJA/action/storage_attestation","attest_author":"https://pith.science/pith/J2K3B3F3SQNZSZFVAOYIMOJPJA/action/author_attestation","sign_citation":"https://pith.science/pith/J2K3B3F3SQNZSZFVAOYIMOJPJA/action/citation_signature","submit_replication":"https://pith.science/pith/J2K3B3F3SQNZSZFVAOYIMOJPJA/action/replication_record"}},"created_at":"2026-07-05T11:01:35.206553+00:00","updated_at":"2026-07-05T11:01:35.206553+00:00"}