{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:MM6TYJA6BMJX4RQMCZTYH2GN7A","short_pith_number":"pith:MM6TYJA6","schema_version":"1.0","canonical_sha256":"633d3c241e0b137e460c166783e8cdf836f617aee92c418f01bcfbd3310c7ca5","source":{"kind":"arxiv","id":"2507.10616","version":2},"attestation_state":"computed","paper":{"title":"Scalpel vs. Hammer: GRPO Amplifies Existing Capabilities, SFT Replaces Them","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Aryo Pradipta Gema, Ivan Titov, Neel Rajani, Seraphina Goldfarb-Tarrant","submitted_at":"2025-07-13T19:04:17Z","abstract_excerpt":"Training large language models (LLMs) for reasoning via maths and code datasets has become a major new focus in LLM post-training. Two particularly popular approaches are reinforcement learning (RL) and supervised fine-tuning (SFT), but their training dynamics are poorly understood. We present a comparative analysis of RL and SFT on the same maths problems with the same model and similar hyperparameters. We find that RL yields minor in-domain gains on maths and slight degradation on knowledge-intensive benchmarks like MMLU, while both trends are more pronounced in SFT. We also analyse model pa"},"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":"2507.10616","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.LG","submitted_at":"2025-07-13T19:04:17Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"1ea180aefd267a8b1dde3f35378b7f05a3d1c26077b219b9bab3db1b80b685bc","abstract_canon_sha256":"167cc0e1f6a7de1d49984d2042c43ba1fb601431a72982f98eb2a1d6c783bef9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:43:13.533549Z","signature_b64":"FR+KMiHXGI292hviDDUJ5b1m88frjEPLai1Dv2z+hXsdD5U9S1iG+K3yc1Aml3iuAkhjRKZ+hvbY97UzlBUQAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"633d3c241e0b137e460c166783e8cdf836f617aee92c418f01bcfbd3310c7ca5","last_reissued_at":"2026-07-05T11:43:13.533083Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:43:13.533083Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Scalpel vs. Hammer: GRPO Amplifies Existing Capabilities, SFT Replaces Them","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Aryo Pradipta Gema, Ivan Titov, Neel Rajani, Seraphina Goldfarb-Tarrant","submitted_at":"2025-07-13T19:04:17Z","abstract_excerpt":"Training large language models (LLMs) for reasoning via maths and code datasets has become a major new focus in LLM post-training. Two particularly popular approaches are reinforcement learning (RL) and supervised fine-tuning (SFT), but their training dynamics are poorly understood. We present a comparative analysis of RL and SFT on the same maths problems with the same model and similar hyperparameters. We find that RL yields minor in-domain gains on maths and slight degradation on knowledge-intensive benchmarks like MMLU, while both trends are more pronounced in SFT. We also analyse model pa"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.10616","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/2507.10616/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":"2507.10616","created_at":"2026-07-05T11:43:13.533134+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.10616v2","created_at":"2026-07-05T11:43:13.533134+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.10616","created_at":"2026-07-05T11:43:13.533134+00:00"},{"alias_kind":"pith_short_12","alias_value":"MM6TYJA6BMJX","created_at":"2026-07-05T11:43:13.533134+00:00"},{"alias_kind":"pith_short_16","alias_value":"MM6TYJA6BMJX4RQM","created_at":"2026-07-05T11:43:13.533134+00:00"},{"alias_kind":"pith_short_8","alias_value":"MM6TYJA6","created_at":"2026-07-05T11:43:13.533134+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2509.25758","citing_title":"Thinking Sparks!: Emergent Attention Heads in Reasoning Models During Post Training","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00610","citing_title":"Decouple before Integration: Test-time Synthesis of SFT and RLVR Task Vectors","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14237","citing_title":"TOPCELL: Topology Optimization of Standard Cell via LLMs","ref_index":27,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MM6TYJA6BMJX4RQMCZTYH2GN7A","json":"https://pith.science/pith/MM6TYJA6BMJX4RQMCZTYH2GN7A.json","graph_json":"https://pith.science/api/pith-number/MM6TYJA6BMJX4RQMCZTYH2GN7A/graph.json","events_json":"https://pith.science/api/pith-number/MM6TYJA6BMJX4RQMCZTYH2GN7A/events.json","paper":"https://pith.science/paper/MM6TYJA6"},"agent_actions":{"view_html":"https://pith.science/pith/MM6TYJA6BMJX4RQMCZTYH2GN7A","download_json":"https://pith.science/pith/MM6TYJA6BMJX4RQMCZTYH2GN7A.json","view_paper":"https://pith.science/paper/MM6TYJA6","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.10616&json=true","fetch_graph":"https://pith.science/api/pith-number/MM6TYJA6BMJX4RQMCZTYH2GN7A/graph.json","fetch_events":"https://pith.science/api/pith-number/MM6TYJA6BMJX4RQMCZTYH2GN7A/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MM6TYJA6BMJX4RQMCZTYH2GN7A/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MM6TYJA6BMJX4RQMCZTYH2GN7A/action/storage_attestation","attest_author":"https://pith.science/pith/MM6TYJA6BMJX4RQMCZTYH2GN7A/action/author_attestation","sign_citation":"https://pith.science/pith/MM6TYJA6BMJX4RQMCZTYH2GN7A/action/citation_signature","submit_replication":"https://pith.science/pith/MM6TYJA6BMJX4RQMCZTYH2GN7A/action/replication_record"}},"created_at":"2026-07-05T11:43:13.533134+00:00","updated_at":"2026-07-05T11:43:13.533134+00:00"}