{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:Z54I7K4FRLCQCXV4SMMFLFTDZ7","short_pith_number":"pith:Z54I7K4F","schema_version":"1.0","canonical_sha256":"cf788fab858ac5015ebc9318559663cfd2141eec5644e2158752bd6d0bc37e2d","source":{"kind":"arxiv","id":"2603.00963","version":2},"attestation_state":"computed","paper":{"title":"Stabilizing Policy Optimization via Logits Convexity","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Hongzhan Chen, Shiping Gao, Tao Yang, Ting Yao, Xiaojun Quan, Yuhua Zhu","submitted_at":"2026-03-01T07:40:12Z","abstract_excerpt":"While reinforcement learning (RL) has been central to the recent success of large language models (LLMs), RL optimization is notoriously unstable, especially when compared to supervised fine-tuning (SFT). In this work, we investigate the stability gap between SFT and RL from a gradient-based perspective, and show that the convexity of the SFT loss with respect to model logits plays a key role in enabling stable training. Our theoretical analysis demonstrates that this property induces favorable gradient directionality during optimization. In contrast, Proximal Policy Optimization (PPO), a wide"},"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":"2603.00963","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-03-01T07:40:12Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"5d388401b511a50152114e4bf1410ebd7a6c81ba09f9ed3b14a0cad9942f3b38","abstract_canon_sha256":"69c137aff0e1e46d6cbba0ab413da813755ab6c966f528c2fc221e0a190c1428"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-02T02:04:15.457242Z","signature_b64":"EorKwjoshu1NHTVJeBDUhONOdvTzRGwPnUsNcTlNxD3ooICsnMxBZLD03W2D5L0IE3hRiWZ0iUcmI7EgN5E5Dw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cf788fab858ac5015ebc9318559663cfd2141eec5644e2158752bd6d0bc37e2d","last_reissued_at":"2026-06-02T02:04:15.456706Z","signature_status":"signed_v1","first_computed_at":"2026-06-02T02:04:15.456706Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Stabilizing Policy Optimization via Logits Convexity","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Hongzhan Chen, Shiping Gao, Tao Yang, Ting Yao, Xiaojun Quan, Yuhua Zhu","submitted_at":"2026-03-01T07:40:12Z","abstract_excerpt":"While reinforcement learning (RL) has been central to the recent success of large language models (LLMs), RL optimization is notoriously unstable, especially when compared to supervised fine-tuning (SFT). In this work, we investigate the stability gap between SFT and RL from a gradient-based perspective, and show that the convexity of the SFT loss with respect to model logits plays a key role in enabling stable training. Our theoretical analysis demonstrates that this property induces favorable gradient directionality during optimization. In contrast, Proximal Policy Optimization (PPO), a wide"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.00963","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/2603.00963/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":"2603.00963","created_at":"2026-06-02T02:04:15.456768+00:00"},{"alias_kind":"arxiv_version","alias_value":"2603.00963v2","created_at":"2026-06-02T02:04:15.456768+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.00963","created_at":"2026-06-02T02:04:15.456768+00:00"},{"alias_kind":"pith_short_12","alias_value":"Z54I7K4FRLCQ","created_at":"2026-06-02T02:04:15.456768+00:00"},{"alias_kind":"pith_short_16","alias_value":"Z54I7K4FRLCQCXV4","created_at":"2026-06-02T02:04:15.456768+00:00"},{"alias_kind":"pith_short_8","alias_value":"Z54I7K4F","created_at":"2026-06-02T02:04:15.456768+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2604.13197","citing_title":"Unleashing Implicit Rewards: Prefix-Value Learning for Distribution-Level Optimization","ref_index":1,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/Z54I7K4FRLCQCXV4SMMFLFTDZ7","json":"https://pith.science/pith/Z54I7K4FRLCQCXV4SMMFLFTDZ7.json","graph_json":"https://pith.science/api/pith-number/Z54I7K4FRLCQCXV4SMMFLFTDZ7/graph.json","events_json":"https://pith.science/api/pith-number/Z54I7K4FRLCQCXV4SMMFLFTDZ7/events.json","paper":"https://pith.science/paper/Z54I7K4F"},"agent_actions":{"view_html":"https://pith.science/pith/Z54I7K4FRLCQCXV4SMMFLFTDZ7","download_json":"https://pith.science/pith/Z54I7K4FRLCQCXV4SMMFLFTDZ7.json","view_paper":"https://pith.science/paper/Z54I7K4F","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2603.00963&json=true","fetch_graph":"https://pith.science/api/pith-number/Z54I7K4FRLCQCXV4SMMFLFTDZ7/graph.json","fetch_events":"https://pith.science/api/pith-number/Z54I7K4FRLCQCXV4SMMFLFTDZ7/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/Z54I7K4FRLCQCXV4SMMFLFTDZ7/action/timestamp_anchor","attest_storage":"https://pith.science/pith/Z54I7K4FRLCQCXV4SMMFLFTDZ7/action/storage_attestation","attest_author":"https://pith.science/pith/Z54I7K4FRLCQCXV4SMMFLFTDZ7/action/author_attestation","sign_citation":"https://pith.science/pith/Z54I7K4FRLCQCXV4SMMFLFTDZ7/action/citation_signature","submit_replication":"https://pith.science/pith/Z54I7K4FRLCQCXV4SMMFLFTDZ7/action/replication_record"}},"created_at":"2026-06-02T02:04:15.456768+00:00","updated_at":"2026-06-02T02:04:15.456768+00:00"}