{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:VUM6XX3CSVZ5NASPUPQO5TLHKU","short_pith_number":"pith:VUM6XX3C","schema_version":"1.0","canonical_sha256":"ad19ebdf629573d6824fa3e0eecd67550ef180bfe9168dd526178e1a1c1b2345","source":{"kind":"arxiv","id":"2506.02553","version":1},"attestation_state":"computed","paper":{"title":"Response-Level Rewards Are All You Need for Online Reinforcement Learning in LLMs: A Mathematical Perspective","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Hui Wei, Shenghua He, Tian Xia, Xuan Zhou","submitted_at":"2025-06-03T07:44:31Z","abstract_excerpt":"We study a common challenge in reinforcement learning for large language models (LLMs): the Zero-Reward Assumption, where non-terminal actions (i.e., intermediate token generations) receive zero task-specific immediate reward, while only the final token receives a reward for the entire response. This assumption arises frequently in practice, as precise token-level rewards are often difficult or infeasible to obtain in LLM applications. In this work, we provide a unifying theoretical perspective. We introduce the Trajectory Policy Gradient Theorem, which shows that the policy gradient based on "},"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":"2506.02553","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-06-03T07:44:31Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"3bd4042b580615d0e709b6ca4fbd4c1d9ca0b79662c4470641071cb80e0ce827","abstract_canon_sha256":"50f6b7bed4f63312b3378a6dd85b3ce2b02f601ed0f782fa1eac3a4a534bb22c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:15:03.942457Z","signature_b64":"dUioU+5TqCEEq1PTxxm/lMbFxYzZMF7l81PLNs6w1F5Oma03nYIUK0UKznpteOBxO/fUVvhdTavhC5ZBq1+LBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ad19ebdf629573d6824fa3e0eecd67550ef180bfe9168dd526178e1a1c1b2345","last_reissued_at":"2026-07-05T11:15:03.941956Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:15:03.941956Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Response-Level Rewards Are All You Need for Online Reinforcement Learning in LLMs: A Mathematical Perspective","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Hui Wei, Shenghua He, Tian Xia, Xuan Zhou","submitted_at":"2025-06-03T07:44:31Z","abstract_excerpt":"We study a common challenge in reinforcement learning for large language models (LLMs): the Zero-Reward Assumption, where non-terminal actions (i.e., intermediate token generations) receive zero task-specific immediate reward, while only the final token receives a reward for the entire response. This assumption arises frequently in practice, as precise token-level rewards are often difficult or infeasible to obtain in LLM applications. In this work, we provide a unifying theoretical perspective. We introduce the Trajectory Policy Gradient Theorem, which shows that the policy gradient based on "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.02553","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/2506.02553/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":"2506.02553","created_at":"2026-07-05T11:15:03.942011+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.02553v1","created_at":"2026-07-05T11:15:03.942011+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.02553","created_at":"2026-07-05T11:15:03.942011+00:00"},{"alias_kind":"pith_short_12","alias_value":"VUM6XX3CSVZ5","created_at":"2026-07-05T11:15:03.942011+00:00"},{"alias_kind":"pith_short_16","alias_value":"VUM6XX3CSVZ5NASP","created_at":"2026-07-05T11:15:03.942011+00:00"},{"alias_kind":"pith_short_8","alias_value":"VUM6XX3C","created_at":"2026-07-05T11:15:03.942011+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2503.09567","citing_title":"Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models","ref_index":252,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10218","citing_title":"Relative Score Policy Optimization for Diffusion Language Models","ref_index":70,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VUM6XX3CSVZ5NASPUPQO5TLHKU","json":"https://pith.science/pith/VUM6XX3CSVZ5NASPUPQO5TLHKU.json","graph_json":"https://pith.science/api/pith-number/VUM6XX3CSVZ5NASPUPQO5TLHKU/graph.json","events_json":"https://pith.science/api/pith-number/VUM6XX3CSVZ5NASPUPQO5TLHKU/events.json","paper":"https://pith.science/paper/VUM6XX3C"},"agent_actions":{"view_html":"https://pith.science/pith/VUM6XX3CSVZ5NASPUPQO5TLHKU","download_json":"https://pith.science/pith/VUM6XX3CSVZ5NASPUPQO5TLHKU.json","view_paper":"https://pith.science/paper/VUM6XX3C","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.02553&json=true","fetch_graph":"https://pith.science/api/pith-number/VUM6XX3CSVZ5NASPUPQO5TLHKU/graph.json","fetch_events":"https://pith.science/api/pith-number/VUM6XX3CSVZ5NASPUPQO5TLHKU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VUM6XX3CSVZ5NASPUPQO5TLHKU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VUM6XX3CSVZ5NASPUPQO5TLHKU/action/storage_attestation","attest_author":"https://pith.science/pith/VUM6XX3CSVZ5NASPUPQO5TLHKU/action/author_attestation","sign_citation":"https://pith.science/pith/VUM6XX3CSVZ5NASPUPQO5TLHKU/action/citation_signature","submit_replication":"https://pith.science/pith/VUM6XX3CSVZ5NASPUPQO5TLHKU/action/replication_record"}},"created_at":"2026-07-05T11:15:03.942011+00:00","updated_at":"2026-07-05T11:15:03.942011+00:00"}