{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:CZSISG5FSUQTO6RP7GPMEE6L4S","short_pith_number":"pith:CZSISG5F","schema_version":"1.0","canonical_sha256":"1664891ba59521377a2ff99ec213cbe49a95e06798b9beba2595e2e319003670","source":{"kind":"arxiv","id":"2510.07364","version":4},"attestation_state":"computed","paper":{"title":"Base Models Know How to Reason, Thinking Models Learn When","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Arthur Conmy, Constantin Venhoff, Iv\\'an Arcuschin, Neel Nanda, Philip Torr","submitted_at":"2025-10-08T17:58:28Z","abstract_excerpt":"What do thinking language models learn during training that their base models lack? We first present an unsupervised method that discovers a model's reasoning behaviors by training small Sparse Autoencoders on sentence-level activations of reasoning traces, yielding interpretable reasoning taxonomies. Building on this, we introduce constructive model diffing, which aims to reconstruct the base-to-fine-tuned difference from interpretable components: reasoning mechanisms (category vectors that can induce a reasoning behavior in the base model) and reasoning heuristics (a classifier determining w"},"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":"2510.07364","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-10-08T17:58:28Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"cccbc72b5fbffa7ca0a8e5427ba42dfa84289c50c99b9ae84ec35364e29fcb01","abstract_canon_sha256":"ee4ca7b7d7de325b2f51c6eef2932eda6e5317cb0a5ef31ac5f130b79f08646b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-08T01:19:07.420685Z","signature_b64":"yvTnTs/uEBGlug1khWVKrRpyfDn4iFiYJ2hXq6UOSIjy4DRfGQrAjLsX7rtn4vzV3AdpLFuilbJ6BQvoxZSLCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1664891ba59521377a2ff99ec213cbe49a95e06798b9beba2595e2e319003670","last_reissued_at":"2026-07-08T01:19:07.420077Z","signature_status":"signed_v1","first_computed_at":"2026-07-08T01:19:07.420077Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Base Models Know How to Reason, Thinking Models Learn When","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Arthur Conmy, Constantin Venhoff, Iv\\'an Arcuschin, Neel Nanda, Philip Torr","submitted_at":"2025-10-08T17:58:28Z","abstract_excerpt":"What do thinking language models learn during training that their base models lack? We first present an unsupervised method that discovers a model's reasoning behaviors by training small Sparse Autoencoders on sentence-level activations of reasoning traces, yielding interpretable reasoning taxonomies. Building on this, we introduce constructive model diffing, which aims to reconstruct the base-to-fine-tuned difference from interpretable components: reasoning mechanisms (category vectors that can induce a reasoning behavior in the base model) and reasoning heuristics (a classifier determining w"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2510.07364","kind":"arxiv","version":4},"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/2510.07364/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":"2510.07364","created_at":"2026-07-08T01:19:07.420147+00:00"},{"alias_kind":"arxiv_version","alias_value":"2510.07364v4","created_at":"2026-07-08T01:19:07.420147+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2510.07364","created_at":"2026-07-08T01:19:07.420147+00:00"},{"alias_kind":"pith_short_12","alias_value":"CZSISG5FSUQT","created_at":"2026-07-08T01:19:07.420147+00:00"},{"alias_kind":"pith_short_16","alias_value":"CZSISG5FSUQTO6RP","created_at":"2026-07-08T01:19:07.420147+00:00"},{"alias_kind":"pith_short_8","alias_value":"CZSISG5F","created_at":"2026-07-08T01:19:07.420147+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":6,"sample":[{"citing_arxiv_id":"2606.18284","citing_title":"Breaking the Solver Bottleneck: Training Task Generators at the Learnable Frontier","ref_index":107,"is_internal_anchor":true},{"citing_arxiv_id":"2603.17839","citing_title":"How do LLMs Compute Verbal Confidence","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2605.17187","citing_title":"PluRule: A Benchmark for Moderating Pluralistic Communities on Social Media","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16874","citing_title":"Reasoning Can Be Restored by Correcting a Few Decision Tokens","ref_index":21,"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":23,"is_internal_anchor":true},{"citing_arxiv_id":"2604.06377","citing_title":"The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment","ref_index":68,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CZSISG5FSUQTO6RP7GPMEE6L4S","json":"https://pith.science/pith/CZSISG5FSUQTO6RP7GPMEE6L4S.json","graph_json":"https://pith.science/api/pith-number/CZSISG5FSUQTO6RP7GPMEE6L4S/graph.json","events_json":"https://pith.science/api/pith-number/CZSISG5FSUQTO6RP7GPMEE6L4S/events.json","paper":"https://pith.science/paper/CZSISG5F"},"agent_actions":{"view_html":"https://pith.science/pith/CZSISG5FSUQTO6RP7GPMEE6L4S","download_json":"https://pith.science/pith/CZSISG5FSUQTO6RP7GPMEE6L4S.json","view_paper":"https://pith.science/paper/CZSISG5F","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2510.07364&json=true","fetch_graph":"https://pith.science/api/pith-number/CZSISG5FSUQTO6RP7GPMEE6L4S/graph.json","fetch_events":"https://pith.science/api/pith-number/CZSISG5FSUQTO6RP7GPMEE6L4S/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CZSISG5FSUQTO6RP7GPMEE6L4S/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CZSISG5FSUQTO6RP7GPMEE6L4S/action/storage_attestation","attest_author":"https://pith.science/pith/CZSISG5FSUQTO6RP7GPMEE6L4S/action/author_attestation","sign_citation":"https://pith.science/pith/CZSISG5FSUQTO6RP7GPMEE6L4S/action/citation_signature","submit_replication":"https://pith.science/pith/CZSISG5FSUQTO6RP7GPMEE6L4S/action/replication_record"}},"created_at":"2026-07-08T01:19:07.420147+00:00","updated_at":"2026-07-08T01:19:07.420147+00:00"}