{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:ERQZIJBHUKS6TY3BH7L56L3ZC6","short_pith_number":"pith:ERQZIJBH","schema_version":"1.0","canonical_sha256":"2461942427a2a5e9e3613fd7df2f79179fe654a7b6c10dae1b3f11f92020dee5","source":{"kind":"arxiv","id":"2410.05459","version":2},"attestation_state":"computed","paper":{"title":"From Sparse Dependence to Sparse Attention: Unveiling How Chain-of-Thought Enhances Transformer Sample Efficiency","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","stat.ML"],"primary_cat":"cs.LG","authors_text":"Hongzhou Lin, Huaqing Zhang, Jingzhao Zhang, Kaiyue Wen","submitted_at":"2024-10-07T19:45:09Z","abstract_excerpt":"Chain-of-thought (CoT) significantly enhances the reasoning performance of large language models (LLM). While current theoretical studies often attribute this improvement to increased expressiveness and computational capacity, we argue that expressiveness is not the primary limitation in the LLM regime, as current large models will fail on simple tasks. Using a parity-learning setup, we demonstrate that CoT can substantially improve sample efficiency even when the representation power is sufficient. Specifically, with CoT, a transformer can learn the function within polynomial samples, whereas"},"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":"2410.05459","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-10-07T19:45:09Z","cross_cats_sorted":["cs.CL","stat.ML"],"title_canon_sha256":"d2e58c13cd729fa3d2505fc9b0fe7a9266ed2c2b8b4624fc974ae7ef8ff22418","abstract_canon_sha256":"f377249166e76366c44895ec04f394ce12fc6fb5f94c13eb5077c05b8a0e4e7f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:24:26.321230Z","signature_b64":"uB4lw3sap5p86axUSARIs2L3cFVOL/C3W7LV6LsMGEBKcldGAbC6MdLjR6NAvJ/dCaAl8NFz3ji+7wCZ4C0AAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2461942427a2a5e9e3613fd7df2f79179fe654a7b6c10dae1b3f11f92020dee5","last_reissued_at":"2026-07-05T10:24:26.320580Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:24:26.320580Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"From Sparse Dependence to Sparse Attention: Unveiling How Chain-of-Thought Enhances Transformer Sample Efficiency","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","stat.ML"],"primary_cat":"cs.LG","authors_text":"Hongzhou Lin, Huaqing Zhang, Jingzhao Zhang, Kaiyue Wen","submitted_at":"2024-10-07T19:45:09Z","abstract_excerpt":"Chain-of-thought (CoT) significantly enhances the reasoning performance of large language models (LLM). While current theoretical studies often attribute this improvement to increased expressiveness and computational capacity, we argue that expressiveness is not the primary limitation in the LLM regime, as current large models will fail on simple tasks. Using a parity-learning setup, we demonstrate that CoT can substantially improve sample efficiency even when the representation power is sufficient. Specifically, with CoT, a transformer can learn the function within polynomial samples, whereas"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.05459","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/2410.05459/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":"2410.05459","created_at":"2026-07-05T10:24:26.320652+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.05459v2","created_at":"2026-07-05T10:24:26.320652+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.05459","created_at":"2026-07-05T10:24:26.320652+00:00"},{"alias_kind":"pith_short_12","alias_value":"ERQZIJBHUKS6","created_at":"2026-07-05T10:24:26.320652+00:00"},{"alias_kind":"pith_short_16","alias_value":"ERQZIJBHUKS6TY3B","created_at":"2026-07-05T10:24:26.320652+00:00"},{"alias_kind":"pith_short_8","alias_value":"ERQZIJBH","created_at":"2026-07-05T10:24:26.320652+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.28600","citing_title":"Transformers Provably Learn to Internalize Chain-of-Thought","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00183","citing_title":"Agentic Transformers Provably Learn to Search via Reinforcement Learning","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23040","citing_title":"Steered Generation via Gradient-Based Optimization on Sparse Query Features","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10019","citing_title":"The two clocks and the innovation window: When and how generative models learn rules","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2604.22951","citing_title":"The Power of Power Law: Asymmetry Enables Compositional Reasoning","ref_index":54,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ERQZIJBHUKS6TY3BH7L56L3ZC6","json":"https://pith.science/pith/ERQZIJBHUKS6TY3BH7L56L3ZC6.json","graph_json":"https://pith.science/api/pith-number/ERQZIJBHUKS6TY3BH7L56L3ZC6/graph.json","events_json":"https://pith.science/api/pith-number/ERQZIJBHUKS6TY3BH7L56L3ZC6/events.json","paper":"https://pith.science/paper/ERQZIJBH"},"agent_actions":{"view_html":"https://pith.science/pith/ERQZIJBHUKS6TY3BH7L56L3ZC6","download_json":"https://pith.science/pith/ERQZIJBHUKS6TY3BH7L56L3ZC6.json","view_paper":"https://pith.science/paper/ERQZIJBH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.05459&json=true","fetch_graph":"https://pith.science/api/pith-number/ERQZIJBHUKS6TY3BH7L56L3ZC6/graph.json","fetch_events":"https://pith.science/api/pith-number/ERQZIJBHUKS6TY3BH7L56L3ZC6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ERQZIJBHUKS6TY3BH7L56L3ZC6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ERQZIJBHUKS6TY3BH7L56L3ZC6/action/storage_attestation","attest_author":"https://pith.science/pith/ERQZIJBHUKS6TY3BH7L56L3ZC6/action/author_attestation","sign_citation":"https://pith.science/pith/ERQZIJBHUKS6TY3BH7L56L3ZC6/action/citation_signature","submit_replication":"https://pith.science/pith/ERQZIJBHUKS6TY3BH7L56L3ZC6/action/replication_record"}},"created_at":"2026-07-05T10:24:26.320652+00:00","updated_at":"2026-07-05T10:24:26.320652+00:00"}