{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:ZP34TZYDFW4BTKWISBW4376LGV","short_pith_number":"pith:ZP34TZYD","schema_version":"1.0","canonical_sha256":"cbf7c9e7032db819aac8906dcdffcb355a8231c9aeeb0dad33792bbcb2f74f40","source":{"kind":"arxiv","id":"2507.22928","version":1},"attestation_state":"computed","paper":{"title":"How does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Aske Plaat, Niki van Stein, Xi Chen","submitted_at":"2025-07-24T10:25:46Z","abstract_excerpt":"Chain-of-thought (CoT) prompting boosts Large Language Models accuracy on multi-step tasks, yet whether the generated \"thoughts\" reflect the true internal reasoning process is unresolved. We present the first feature-level causal study of CoT faithfulness. Combining sparse autoencoders with activation patching, we extract monosemantic features from Pythia-70M and Pythia-2.8B while they tackle GSM8K math problems under CoT and plain (noCoT) prompting. Swapping a small set of CoT-reasoning features into a noCoT run raises answer log-probabilities significantly in the 2.8B model, but has no relia"},"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":"2507.22928","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-07-24T10:25:46Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"cc569f6515e502fc17e805cb5780b577b4ee36d9a3f4c437aa8754300718a51e","abstract_canon_sha256":"6cea1fc211cbd377802526c6d48f68e0729b878cd0346e30db928e546e7109da"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:46:02.096960Z","signature_b64":"Vwk8jKb/EL8f6eNoma9ueddSUTel0QE6YZw2ZLGopX1kGhr8MtDZKsLgVMhYLaeq/f7qaj/+dEr6Vl5paC4WAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cbf7c9e7032db819aac8906dcdffcb355a8231c9aeeb0dad33792bbcb2f74f40","last_reissued_at":"2026-07-05T11:46:02.096466Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:46:02.096466Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"How does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Aske Plaat, Niki van Stein, Xi Chen","submitted_at":"2025-07-24T10:25:46Z","abstract_excerpt":"Chain-of-thought (CoT) prompting boosts Large Language Models accuracy on multi-step tasks, yet whether the generated \"thoughts\" reflect the true internal reasoning process is unresolved. We present the first feature-level causal study of CoT faithfulness. Combining sparse autoencoders with activation patching, we extract monosemantic features from Pythia-70M and Pythia-2.8B while they tackle GSM8K math problems under CoT and plain (noCoT) prompting. Swapping a small set of CoT-reasoning features into a noCoT run raises answer log-probabilities significantly in the 2.8B model, but has no relia"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.22928","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/2507.22928/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":"2507.22928","created_at":"2026-07-05T11:46:02.096528+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.22928v1","created_at":"2026-07-05T11:46:02.096528+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.22928","created_at":"2026-07-05T11:46:02.096528+00:00"},{"alias_kind":"pith_short_12","alias_value":"ZP34TZYDFW4B","created_at":"2026-07-05T11:46:02.096528+00:00"},{"alias_kind":"pith_short_16","alias_value":"ZP34TZYDFW4BTKWI","created_at":"2026-07-05T11:46:02.096528+00:00"},{"alias_kind":"pith_short_8","alias_value":"ZP34TZYD","created_at":"2026-07-05T11:46:02.096528+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.06188","citing_title":"The Tell-Tale Norm: $\\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06840","citing_title":"Characterize Then Distill: Mechanistic Reasoning in Large Output Spaces","ref_index":88,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ZP34TZYDFW4BTKWISBW4376LGV","json":"https://pith.science/pith/ZP34TZYDFW4BTKWISBW4376LGV.json","graph_json":"https://pith.science/api/pith-number/ZP34TZYDFW4BTKWISBW4376LGV/graph.json","events_json":"https://pith.science/api/pith-number/ZP34TZYDFW4BTKWISBW4376LGV/events.json","paper":"https://pith.science/paper/ZP34TZYD"},"agent_actions":{"view_html":"https://pith.science/pith/ZP34TZYDFW4BTKWISBW4376LGV","download_json":"https://pith.science/pith/ZP34TZYDFW4BTKWISBW4376LGV.json","view_paper":"https://pith.science/paper/ZP34TZYD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.22928&json=true","fetch_graph":"https://pith.science/api/pith-number/ZP34TZYDFW4BTKWISBW4376LGV/graph.json","fetch_events":"https://pith.science/api/pith-number/ZP34TZYDFW4BTKWISBW4376LGV/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ZP34TZYDFW4BTKWISBW4376LGV/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ZP34TZYDFW4BTKWISBW4376LGV/action/storage_attestation","attest_author":"https://pith.science/pith/ZP34TZYDFW4BTKWISBW4376LGV/action/author_attestation","sign_citation":"https://pith.science/pith/ZP34TZYDFW4BTKWISBW4376LGV/action/citation_signature","submit_replication":"https://pith.science/pith/ZP34TZYDFW4BTKWISBW4376LGV/action/replication_record"}},"created_at":"2026-07-05T11:46:02.096528+00:00","updated_at":"2026-07-05T11:46:02.096528+00:00"}