{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:ZHIT4CRQBENNHMFUHHASST3YGH","short_pith_number":"pith:ZHIT4CRQ","schema_version":"1.0","canonical_sha256":"c9d13e0a30091ad3b0b439c1294f7831f543d752e6e03acee3e73109bfb98569","source":{"kind":"arxiv","id":"2509.21128","version":2},"attestation_state":"computed","paper":{"title":"RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Gouki Minegishi, Kohsei Matsutani, Shota Takashiro, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo","submitted_at":"2025-09-25T13:18:57Z","abstract_excerpt":"Large language models (LLMs) are typically trained by reinforcement learning (RL) with verifiable rewards (RLVR) and supervised fine-tuning (SFT) on reasoning traces to improve their reasoning abilities. However, how these methods shape reasoning capabilities remains largely elusive. Going beyond an accuracy-based investigation of how these two components sculpt the reasoning process, this paper introduces a novel analysis framework that quantifies reasoning paths and captures their qualitative changes under each training process (with models of 1.5B, 7B, and 14B parameters on mathematical dom"},"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":"2509.21128","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-09-25T13:18:57Z","cross_cats_sorted":[],"title_canon_sha256":"c8254ca837dad65a9e2446188454fb13e985b67294beb6fd7c9a4f24092a0495","abstract_canon_sha256":"637ca2d069157c3799d74ddb9c46057e66d1bd63915e349a1fa72e25519b9e46"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-28T01:04:30.716116Z","signature_b64":"N5Fo1vl184H/COOrGU0Bqm5CGmjUjqMCr+kFwyZQfxgt1B8bjdDeoz0iKPA2Wm7N45DG9gPcvwO/XAJfzSoVDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c9d13e0a30091ad3b0b439c1294f7831f543d752e6e03acee3e73109bfb98569","last_reissued_at":"2026-05-28T01:04:30.715572Z","signature_status":"signed_v1","first_computed_at":"2026-05-28T01:04:30.715572Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Gouki Minegishi, Kohsei Matsutani, Shota Takashiro, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo","submitted_at":"2025-09-25T13:18:57Z","abstract_excerpt":"Large language models (LLMs) are typically trained by reinforcement learning (RL) with verifiable rewards (RLVR) and supervised fine-tuning (SFT) on reasoning traces to improve their reasoning abilities. However, how these methods shape reasoning capabilities remains largely elusive. Going beyond an accuracy-based investigation of how these two components sculpt the reasoning process, this paper introduces a novel analysis framework that quantifies reasoning paths and captures their qualitative changes under each training process (with models of 1.5B, 7B, and 14B parameters on mathematical dom"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2509.21128","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/2509.21128/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":"2509.21128","created_at":"2026-05-28T01:04:30.715645+00:00"},{"alias_kind":"arxiv_version","alias_value":"2509.21128v2","created_at":"2026-05-28T01:04:30.715645+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2509.21128","created_at":"2026-05-28T01:04:30.715645+00:00"},{"alias_kind":"pith_short_12","alias_value":"ZHIT4CRQBENN","created_at":"2026-05-28T01:04:30.715645+00:00"},{"alias_kind":"pith_short_16","alias_value":"ZHIT4CRQBENNHMFU","created_at":"2026-05-28T01:04:30.715645+00:00"},{"alias_kind":"pith_short_8","alias_value":"ZHIT4CRQ","created_at":"2026-05-28T01:04:30.715645+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":4,"sample":[{"citing_arxiv_id":"2605.13080","citing_title":"Learning to See What You Need: Gaze Attention for Multimodal Large Language Models","ref_index":111,"is_internal_anchor":true},{"citing_arxiv_id":"2605.00610","citing_title":"Decouple before Integration: Test-time Synthesis of SFT and RLVR Task Vectors","ref_index":45,"is_internal_anchor":true},{"citing_arxiv_id":"2605.00365","citing_title":"Uniform-Correct Policy Optimization: Breaking RLVR's Indifference to Diversity","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2603.01865","citing_title":"CyclicJudge: Mitigating Judge Bias Efficiently in LLM-based Evaluation","ref_index":3,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ZHIT4CRQBENNHMFUHHASST3YGH","json":"https://pith.science/pith/ZHIT4CRQBENNHMFUHHASST3YGH.json","graph_json":"https://pith.science/api/pith-number/ZHIT4CRQBENNHMFUHHASST3YGH/graph.json","events_json":"https://pith.science/api/pith-number/ZHIT4CRQBENNHMFUHHASST3YGH/events.json","paper":"https://pith.science/paper/ZHIT4CRQ"},"agent_actions":{"view_html":"https://pith.science/pith/ZHIT4CRQBENNHMFUHHASST3YGH","download_json":"https://pith.science/pith/ZHIT4CRQBENNHMFUHHASST3YGH.json","view_paper":"https://pith.science/paper/ZHIT4CRQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2509.21128&json=true","fetch_graph":"https://pith.science/api/pith-number/ZHIT4CRQBENNHMFUHHASST3YGH/graph.json","fetch_events":"https://pith.science/api/pith-number/ZHIT4CRQBENNHMFUHHASST3YGH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ZHIT4CRQBENNHMFUHHASST3YGH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ZHIT4CRQBENNHMFUHHASST3YGH/action/storage_attestation","attest_author":"https://pith.science/pith/ZHIT4CRQBENNHMFUHHASST3YGH/action/author_attestation","sign_citation":"https://pith.science/pith/ZHIT4CRQBENNHMFUHHASST3YGH/action/citation_signature","submit_replication":"https://pith.science/pith/ZHIT4CRQBENNHMFUHHASST3YGH/action/replication_record"}},"created_at":"2026-05-28T01:04:30.715645+00:00","updated_at":"2026-05-28T01:04:30.715645+00:00"}