{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:NVLQ2IZZF3HHL4EC5NFV3WQWVN","short_pith_number":"pith:NVLQ2IZZ","schema_version":"1.0","canonical_sha256":"6d570d23392ece75f082eb4b5dda16ab40f5c8d6bf033a864812b186bb4d82a0","source":{"kind":"arxiv","id":"2507.08267","version":1},"attestation_state":"computed","paper":{"title":"A Practical Two-Stage Recipe for Mathematical LLMs: Maximizing Accuracy with SFT and Efficiency with Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Hiroshi Yoshihara, Taiki Yamaguchi, Yuichi Inoue","submitted_at":"2025-07-11T02:26:01Z","abstract_excerpt":"Enhancing the mathematical reasoning of Large Language Models (LLMs) is a pivotal challenge in advancing AI capabilities. While Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are the dominant training paradigms, a systematic methodology for combining them to maximize both accuracy and efficiency remains largely unexplored. This paper introduces a practical and effective training recipe that strategically integrates extended SFT with RL from online inference (GRPO). We posit that these methods play complementary, not competing, roles: a prolonged SFT phase first pushes the model's"},"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":"2507.08267","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-07-11T02:26:01Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"bddc4cb7f5c6f8a1fb9691c3099e806c55be529f38019fd8a4739f7f282ca5aa","abstract_canon_sha256":"36280b93c40846fb0887a1e44aca7cd548b00fda09e3de0eff3e12471c6862fb"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:35:20.361427Z","signature_b64":"r3hi5pno1eAcj0AhHQBpMycIY0XHq3M3fgh1qm6aR/tM3OrMF2TsbMomm8X8Gxy9CtNnAAElWpBVLgj0cPLfBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6d570d23392ece75f082eb4b5dda16ab40f5c8d6bf033a864812b186bb4d82a0","last_reissued_at":"2026-07-05T11:35:20.360835Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:35:20.360835Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Practical Two-Stage Recipe for Mathematical LLMs: Maximizing Accuracy with SFT and Efficiency with Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Hiroshi Yoshihara, Taiki Yamaguchi, Yuichi Inoue","submitted_at":"2025-07-11T02:26:01Z","abstract_excerpt":"Enhancing the mathematical reasoning of Large Language Models (LLMs) is a pivotal challenge in advancing AI capabilities. While Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are the dominant training paradigms, a systematic methodology for combining them to maximize both accuracy and efficiency remains largely unexplored. This paper introduces a practical and effective training recipe that strategically integrates extended SFT with RL from online inference (GRPO). We posit that these methods play complementary, not competing, roles: a prolonged SFT phase first pushes the model's"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.08267","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/2507.08267/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":"2507.08267","created_at":"2026-07-05T11:35:20.360897+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.08267v1","created_at":"2026-07-05T11:35:20.360897+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.08267","created_at":"2026-07-05T11:35:20.360897+00:00"},{"alias_kind":"pith_short_12","alias_value":"NVLQ2IZZF3HH","created_at":"2026-07-05T11:35:20.360897+00:00"},{"alias_kind":"pith_short_16","alias_value":"NVLQ2IZZF3HHL4EC","created_at":"2026-07-05T11:35:20.360897+00:00"},{"alias_kind":"pith_short_8","alias_value":"NVLQ2IZZ","created_at":"2026-07-05T11:35:20.360897+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2512.11470","citing_title":"Rethinking Expert Trajectory Utilization in LLM Post-training for Mathematical Reasoning","ref_index":47,"is_internal_anchor":false},{"citing_arxiv_id":"2603.11321","citing_title":"Hindsight-Anchored Policy Optimization: Turning Failure into Feedback in Sparse Reward Settings","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08926","citing_title":"Bridging SFT and RL: Dynamic Policy Optimization for Robust Reasoning","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06079","citing_title":"Scientific Graphics Program Synthesis via Dual Self-Consistency Reinforcement Learning","ref_index":52,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NVLQ2IZZF3HHL4EC5NFV3WQWVN","json":"https://pith.science/pith/NVLQ2IZZF3HHL4EC5NFV3WQWVN.json","graph_json":"https://pith.science/api/pith-number/NVLQ2IZZF3HHL4EC5NFV3WQWVN/graph.json","events_json":"https://pith.science/api/pith-number/NVLQ2IZZF3HHL4EC5NFV3WQWVN/events.json","paper":"https://pith.science/paper/NVLQ2IZZ"},"agent_actions":{"view_html":"https://pith.science/pith/NVLQ2IZZF3HHL4EC5NFV3WQWVN","download_json":"https://pith.science/pith/NVLQ2IZZF3HHL4EC5NFV3WQWVN.json","view_paper":"https://pith.science/paper/NVLQ2IZZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.08267&json=true","fetch_graph":"https://pith.science/api/pith-number/NVLQ2IZZF3HHL4EC5NFV3WQWVN/graph.json","fetch_events":"https://pith.science/api/pith-number/NVLQ2IZZF3HHL4EC5NFV3WQWVN/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NVLQ2IZZF3HHL4EC5NFV3WQWVN/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NVLQ2IZZF3HHL4EC5NFV3WQWVN/action/storage_attestation","attest_author":"https://pith.science/pith/NVLQ2IZZF3HHL4EC5NFV3WQWVN/action/author_attestation","sign_citation":"https://pith.science/pith/NVLQ2IZZF3HHL4EC5NFV3WQWVN/action/citation_signature","submit_replication":"https://pith.science/pith/NVLQ2IZZF3HHL4EC5NFV3WQWVN/action/replication_record"}},"created_at":"2026-07-05T11:35:20.360897+00:00","updated_at":"2026-07-05T11:35:20.360897+00:00"}