{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:MPY37EC35PND67VPTUSH5VMPIM","short_pith_number":"pith:MPY37EC3","schema_version":"1.0","canonical_sha256":"63f1bf905bebda3f7eaf9d247ed58f431fb95d4fa470ae00a8fa736e9007e7d1","source":{"kind":"arxiv","id":"2601.10201","version":2},"attestation_state":"computed","paper":{"title":"Future-KL Regularized GRPO: Process-Level Credit Assignment from $f$-Divergence Regularization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Hao Bai, Jiarui Yao, Ruida Wang, Tong Zhang","submitted_at":"2026-01-15T09:01:53Z","abstract_excerpt":"Group Relative Policy Optimization (GRPO) is widely used for critic-free Large Language Model (LLM) post-training, but its KL regularization is usually implemented as a local loss-side token penalty. We show that this misses the policy-gradient signal induced by autoregressive KL regularization. Unlike standard KL-regularized Reinforcement Learning (RL) objectives, GRPO's group normalization induces a non-linear prompt-level utility; for binary verifier rewards, this utility is $2\\arcsin\\sqrt p$. As a result, reward and KL cannot be fused before normalization without changing the implicit obje"},"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":"2601.10201","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-01-15T09:01:53Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"3b56633fcee5ca25fdef4a5bd86f5f6bcde161572263b70e7cf7502e44d27478","abstract_canon_sha256":"11c63165d4d01f90cdbd20e62c85b206af88476be812139456a5e964bf6c117a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-26T01:02:30.736039Z","signature_b64":"pKfhgs6mKd/wH8dgq7zdTtz4CNAG3OEXjFMDKBJb1K1dADa8NtHnEnpQeOes0Qno+l4pTJxfjmE0MRdx3IEtAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"63f1bf905bebda3f7eaf9d247ed58f431fb95d4fa470ae00a8fa736e9007e7d1","last_reissued_at":"2026-05-26T01:02:30.735217Z","signature_status":"signed_v1","first_computed_at":"2026-05-26T01:02:30.735217Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Future-KL Regularized GRPO: Process-Level Credit Assignment from $f$-Divergence Regularization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Hao Bai, Jiarui Yao, Ruida Wang, Tong Zhang","submitted_at":"2026-01-15T09:01:53Z","abstract_excerpt":"Group Relative Policy Optimization (GRPO) is widely used for critic-free Large Language Model (LLM) post-training, but its KL regularization is usually implemented as a local loss-side token penalty. We show that this misses the policy-gradient signal induced by autoregressive KL regularization. Unlike standard KL-regularized Reinforcement Learning (RL) objectives, GRPO's group normalization induces a non-linear prompt-level utility; for binary verifier rewards, this utility is $2\\arcsin\\sqrt p$. As a result, reward and KL cannot be fused before normalization without changing the implicit obje"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2601.10201","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/2601.10201/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":"2601.10201","created_at":"2026-05-26T01:02:30.735332+00:00"},{"alias_kind":"arxiv_version","alias_value":"2601.10201v2","created_at":"2026-05-26T01:02:30.735332+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.10201","created_at":"2026-05-26T01:02:30.735332+00:00"},{"alias_kind":"pith_short_12","alias_value":"MPY37EC35PND","created_at":"2026-05-26T01:02:30.735332+00:00"},{"alias_kind":"pith_short_16","alias_value":"MPY37EC35PND67VP","created_at":"2026-05-26T01:02:30.735332+00:00"},{"alias_kind":"pith_short_8","alias_value":"MPY37EC3","created_at":"2026-05-26T01:02:30.735332+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":4,"sample":[{"citing_arxiv_id":"2604.02341","citing_title":"LLM Reasoning with Process Rewards for Outcome-Guided Steps","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10158","citing_title":"Unsupervised Process Reward Models","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2604.09459","citing_title":"From Reasoning to Agentic: Credit Assignment in Reinforcement Learning for Large Language Models","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05226","citing_title":"Internalizing Outcome Supervision into Process Supervision: A New Paradigm for Reinforcement Learning for Reasoning","ref_index":26,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MPY37EC35PND67VPTUSH5VMPIM","json":"https://pith.science/pith/MPY37EC35PND67VPTUSH5VMPIM.json","graph_json":"https://pith.science/api/pith-number/MPY37EC35PND67VPTUSH5VMPIM/graph.json","events_json":"https://pith.science/api/pith-number/MPY37EC35PND67VPTUSH5VMPIM/events.json","paper":"https://pith.science/paper/MPY37EC3"},"agent_actions":{"view_html":"https://pith.science/pith/MPY37EC35PND67VPTUSH5VMPIM","download_json":"https://pith.science/pith/MPY37EC35PND67VPTUSH5VMPIM.json","view_paper":"https://pith.science/paper/MPY37EC3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2601.10201&json=true","fetch_graph":"https://pith.science/api/pith-number/MPY37EC35PND67VPTUSH5VMPIM/graph.json","fetch_events":"https://pith.science/api/pith-number/MPY37EC35PND67VPTUSH5VMPIM/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MPY37EC35PND67VPTUSH5VMPIM/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MPY37EC35PND67VPTUSH5VMPIM/action/storage_attestation","attest_author":"https://pith.science/pith/MPY37EC35PND67VPTUSH5VMPIM/action/author_attestation","sign_citation":"https://pith.science/pith/MPY37EC35PND67VPTUSH5VMPIM/action/citation_signature","submit_replication":"https://pith.science/pith/MPY37EC35PND67VPTUSH5VMPIM/action/replication_record"}},"created_at":"2026-05-26T01:02:30.735332+00:00","updated_at":"2026-05-26T01:02:30.735332+00:00"}