{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:KLUVLCAPMUY6IM2QWA2G6AXEBD","short_pith_number":"pith:KLUVLCAP","schema_version":"1.0","canonical_sha256":"52e955880f6531e43350b0346f02e408c1bf67261e55513b114bcf28a75440d9","source":{"kind":"arxiv","id":"2502.04375","version":2},"attestation_state":"computed","paper":{"title":"An Analysis for Reasoning Bias of Language Models with Small Initialization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Junjie Yao, Zhi-Qin John Xu, Zhongwang Zhang","submitted_at":"2025-02-05T15:23:26Z","abstract_excerpt":"Transformer-based Large Language Models (LLMs) have revolutionized Natural Language Processing by demonstrating exceptional performance across diverse tasks. This study investigates the impact of the parameter initialization scale on the training behavior and task preferences of LLMs. We discover that smaller initialization scales encourage models to favor reasoning tasks, whereas larger initialization scales lead to a preference for memorization tasks. We validate this reasoning bias via real datasets and meticulously designed anchor functions. Further analysis of initial training dynamics su"},"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":"2502.04375","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-02-05T15:23:26Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"daefa28694852db56040b92ec14ba8592184dbf6237182bc5ff8293baa3228e7","abstract_canon_sha256":"60e7b326981c8ae47361c7b2e40f449069381dad91e4582ed4699cc89b09ddde"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:06:27.830642Z","signature_b64":"wpvROwWwOI//w7vN1cBgcaatlKYHFx+lwmkYL1XUfg6tfbVuX/5UxZacvNb/YI+IbN/8ybUwSXyF8IqqRnggAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"52e955880f6531e43350b0346f02e408c1bf67261e55513b114bcf28a75440d9","last_reissued_at":"2026-07-05T11:06:27.830162Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:06:27.830162Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"An Analysis for Reasoning Bias of Language Models with Small Initialization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Junjie Yao, Zhi-Qin John Xu, Zhongwang Zhang","submitted_at":"2025-02-05T15:23:26Z","abstract_excerpt":"Transformer-based Large Language Models (LLMs) have revolutionized Natural Language Processing by demonstrating exceptional performance across diverse tasks. This study investigates the impact of the parameter initialization scale on the training behavior and task preferences of LLMs. We discover that smaller initialization scales encourage models to favor reasoning tasks, whereas larger initialization scales lead to a preference for memorization tasks. We validate this reasoning bias via real datasets and meticulously designed anchor functions. Further analysis of initial training dynamics su"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.04375","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/2502.04375/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":"2502.04375","created_at":"2026-07-05T11:06:27.830210+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.04375v2","created_at":"2026-07-05T11:06:27.830210+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.04375","created_at":"2026-07-05T11:06:27.830210+00:00"},{"alias_kind":"pith_short_12","alias_value":"KLUVLCAPMUY6","created_at":"2026-07-05T11:06:27.830210+00:00"},{"alias_kind":"pith_short_16","alias_value":"KLUVLCAPMUY6IM2Q","created_at":"2026-07-05T11:06:27.830210+00:00"},{"alias_kind":"pith_short_8","alias_value":"KLUVLCAP","created_at":"2026-07-05T11:06:27.830210+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2504.09484","citing_title":"An overview of condensation phenomenon in deep learning","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2601.19208","citing_title":"How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2603.01097","citing_title":"Understanding LoRA as Knowledge Memory: An Empirical Analysis","ref_index":3,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/KLUVLCAPMUY6IM2QWA2G6AXEBD","json":"https://pith.science/pith/KLUVLCAPMUY6IM2QWA2G6AXEBD.json","graph_json":"https://pith.science/api/pith-number/KLUVLCAPMUY6IM2QWA2G6AXEBD/graph.json","events_json":"https://pith.science/api/pith-number/KLUVLCAPMUY6IM2QWA2G6AXEBD/events.json","paper":"https://pith.science/paper/KLUVLCAP"},"agent_actions":{"view_html":"https://pith.science/pith/KLUVLCAPMUY6IM2QWA2G6AXEBD","download_json":"https://pith.science/pith/KLUVLCAPMUY6IM2QWA2G6AXEBD.json","view_paper":"https://pith.science/paper/KLUVLCAP","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.04375&json=true","fetch_graph":"https://pith.science/api/pith-number/KLUVLCAPMUY6IM2QWA2G6AXEBD/graph.json","fetch_events":"https://pith.science/api/pith-number/KLUVLCAPMUY6IM2QWA2G6AXEBD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/KLUVLCAPMUY6IM2QWA2G6AXEBD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/KLUVLCAPMUY6IM2QWA2G6AXEBD/action/storage_attestation","attest_author":"https://pith.science/pith/KLUVLCAPMUY6IM2QWA2G6AXEBD/action/author_attestation","sign_citation":"https://pith.science/pith/KLUVLCAPMUY6IM2QWA2G6AXEBD/action/citation_signature","submit_replication":"https://pith.science/pith/KLUVLCAPMUY6IM2QWA2G6AXEBD/action/replication_record"}},"created_at":"2026-07-05T11:06:27.830210+00:00","updated_at":"2026-07-05T11:06:27.830210+00:00"}