{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:FGDSKN473HL4NEH7AXDK4NEUEX","short_pith_number":"pith:FGDSKN47","schema_version":"1.0","canonical_sha256":"298725379fd9d7c690ff05c6ae349425de5de04c0ba5aee7e0b3579ded2d0b5c","source":{"kind":"arxiv","id":"2508.16546","version":1},"attestation_state":"computed","paper":{"title":"RL Is Neither a Panacea Nor a Mirage: Understanding Supervised vs. Reinforcement Learning Fine-Tuning for LLMs","license":"http://creativecommons.org/publicdomain/zero/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Hangzhan Jin, Mohammad Hamdaqa, Sicheng Lv, Sifan Wu","submitted_at":"2025-08-22T17:10:37Z","abstract_excerpt":"Training large language models (LLMs) from scratch is increasingly impractical, making post-training methods such as supervised fine-tuning (SFT) and reinforcement-learning fine-tuning (RL-FT, e.g., PPO) central to modern practice. Using an out-of-distribution (OOD) variant of the 24-point card game and new spectrum-based diagnostics, we revisit how these two stages reshape model representation and OOD performance. Our key findings are- (1) RL-FT can restore much of the OOD performance loss from SFT (e.g., Llama-11B 8.97% to 15.38%, Qwen-7B 17.09% to 19.66%). But when SFT induces severe overfi"},"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":"2508.16546","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/publicdomain/zero/1.0/","primary_cat":"cs.LG","submitted_at":"2025-08-22T17:10:37Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"297c7addee0d1385b07e4805b92aac0b21a0d812bed4193a26c64ed4fb850666","abstract_canon_sha256":"148befcaa82b8769f0259cf1cbb242e3c70548ed9792aff45a0ca7be782771a6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:57:50.531154Z","signature_b64":"bQ7b4gSj4OWb8vpr0czhlweR58lHmq1whuqXrfo8fH/O0cX5Im4a/6NAjI0C58W7TXx3KZxEuTLJk+bfN3aCDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"298725379fd9d7c690ff05c6ae349425de5de04c0ba5aee7e0b3579ded2d0b5c","last_reissued_at":"2026-07-05T11:57:50.530665Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:57:50.530665Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"RL Is Neither a Panacea Nor a Mirage: Understanding Supervised vs. Reinforcement Learning Fine-Tuning for LLMs","license":"http://creativecommons.org/publicdomain/zero/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Hangzhan Jin, Mohammad Hamdaqa, Sicheng Lv, Sifan Wu","submitted_at":"2025-08-22T17:10:37Z","abstract_excerpt":"Training large language models (LLMs) from scratch is increasingly impractical, making post-training methods such as supervised fine-tuning (SFT) and reinforcement-learning fine-tuning (RL-FT, e.g., PPO) central to modern practice. Using an out-of-distribution (OOD) variant of the 24-point card game and new spectrum-based diagnostics, we revisit how these two stages reshape model representation and OOD performance. Our key findings are- (1) RL-FT can restore much of the OOD performance loss from SFT (e.g., Llama-11B 8.97% to 15.38%, Qwen-7B 17.09% to 19.66%). But when SFT induces severe overfi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2508.16546","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/2508.16546/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":"2508.16546","created_at":"2026-07-05T11:57:50.530728+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.16546v1","created_at":"2026-07-05T11:57:50.530728+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.16546","created_at":"2026-07-05T11:57:50.530728+00:00"},{"alias_kind":"pith_short_12","alias_value":"FGDSKN473HL4","created_at":"2026-07-05T11:57:50.530728+00:00"},{"alias_kind":"pith_short_16","alias_value":"FGDSKN473HL4NEH7","created_at":"2026-07-05T11:57:50.530728+00:00"},{"alias_kind":"pith_short_8","alias_value":"FGDSKN47","created_at":"2026-07-05T11:57:50.530728+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.18521","citing_title":"Sparsity Curse: Understanding RLVR Model Parameter Space from Model Merging","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09932","citing_title":"When RL Fails after SFT: Rejuvenating Model Plasticity for Robust SFT-to-RL Handoff","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2509.08827","citing_title":"A Survey of Reinforcement Learning for Large Reasoning Models","ref_index":247,"is_internal_anchor":false},{"citing_arxiv_id":"2604.19945","citing_title":"Visual Reasoning through Tool-supervised Reinforcement Learning","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06628","citing_title":"Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability","ref_index":2,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/FGDSKN473HL4NEH7AXDK4NEUEX","json":"https://pith.science/pith/FGDSKN473HL4NEH7AXDK4NEUEX.json","graph_json":"https://pith.science/api/pith-number/FGDSKN473HL4NEH7AXDK4NEUEX/graph.json","events_json":"https://pith.science/api/pith-number/FGDSKN473HL4NEH7AXDK4NEUEX/events.json","paper":"https://pith.science/paper/FGDSKN47"},"agent_actions":{"view_html":"https://pith.science/pith/FGDSKN473HL4NEH7AXDK4NEUEX","download_json":"https://pith.science/pith/FGDSKN473HL4NEH7AXDK4NEUEX.json","view_paper":"https://pith.science/paper/FGDSKN47","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.16546&json=true","fetch_graph":"https://pith.science/api/pith-number/FGDSKN473HL4NEH7AXDK4NEUEX/graph.json","fetch_events":"https://pith.science/api/pith-number/FGDSKN473HL4NEH7AXDK4NEUEX/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FGDSKN473HL4NEH7AXDK4NEUEX/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FGDSKN473HL4NEH7AXDK4NEUEX/action/storage_attestation","attest_author":"https://pith.science/pith/FGDSKN473HL4NEH7AXDK4NEUEX/action/author_attestation","sign_citation":"https://pith.science/pith/FGDSKN473HL4NEH7AXDK4NEUEX/action/citation_signature","submit_replication":"https://pith.science/pith/FGDSKN473HL4NEH7AXDK4NEUEX/action/replication_record"}},"created_at":"2026-07-05T11:57:50.530728+00:00","updated_at":"2026-07-05T11:57:50.530728+00:00"}