{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:OHHQ2XTR6254YJFUOZYXM4TFND","short_pith_number":"pith:OHHQ2XTR","schema_version":"1.0","canonical_sha256":"71cf0d5e71f6bbcc24b4767176726568fb0f8d0a8031e731fa8e54c4e1654ca9","source":{"kind":"arxiv","id":"2510.00977","version":3},"attestation_state":"computed","paper":{"title":"It Takes Two: Your GRPO Is Secretly DPO","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Chenyang Huang, Jian-Yun Nie, Kejia Chen, Lei Ding, Liheng Ma, Mark Coates, Muzhi Li, Xinyu Wang, Yihong Wu, Yingxue Zhang, Zhanguang Zhang, Zhan Su","submitted_at":"2025-10-01T14:52:11Z","abstract_excerpt":"GRPO has emerged as a prominent reinforcement learning algorithm for post-training LLMs. Unlike critic-based methods, GRPO computes advantages by estimating the \\emph{value baselines} from group-level statistics, eliminating the need for a critic network. Consequently, the prevailing view emphasizes the necessity of large group sizes, which are assumed to yield more accurate statistical estimates. In this paper, we propose a different view that the efficacy of GRPO stems from its implicit contrastive objective in the optimization, which helps reduce variance via the control variate method. Thi"},"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":"2510.00977","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2025-10-01T14:52:11Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"8b357b7795e76241955f62a7d7b64b7329669151c11d1fe1851be1d2c5a2b2ca","abstract_canon_sha256":"a537643ed34f35265a8ce2d9ad6a1e649e6c0b1c0fa78d3f19951782d0e83e53"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:17.355952Z","signature_b64":"yOaL+GEznuu2jmH03hSy0T/445oXqcFk0paUdLjeQ+3lvHU3wpjoiwRGJGrzqKPllSem/qLOQiugi1G5BvAxCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"71cf0d5e71f6bbcc24b4767176726568fb0f8d0a8031e731fa8e54c4e1654ca9","last_reissued_at":"2026-05-17T23:39:17.355423Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:17.355423Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"It Takes Two: Your GRPO Is Secretly DPO","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Chenyang Huang, Jian-Yun Nie, Kejia Chen, Lei Ding, Liheng Ma, Mark Coates, Muzhi Li, Xinyu Wang, Yihong Wu, Yingxue Zhang, Zhanguang Zhang, Zhan Su","submitted_at":"2025-10-01T14:52:11Z","abstract_excerpt":"GRPO has emerged as a prominent reinforcement learning algorithm for post-training LLMs. Unlike critic-based methods, GRPO computes advantages by estimating the \\emph{value baselines} from group-level statistics, eliminating the need for a critic network. Consequently, the prevailing view emphasizes the necessity of large group sizes, which are assumed to yield more accurate statistical estimates. In this paper, we propose a different view that the efficacy of GRPO stems from its implicit contrastive objective in the optimization, which helps reduce variance via the control variate method. Thi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2510.00977","kind":"arxiv","version":3},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"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":"2510.00977","created_at":"2026-05-17T23:39:17.355495+00:00"},{"alias_kind":"arxiv_version","alias_value":"2510.00977v3","created_at":"2026-05-17T23:39:17.355495+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2510.00977","created_at":"2026-05-17T23:39:17.355495+00:00"},{"alias_kind":"pith_short_12","alias_value":"OHHQ2XTR6254","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"OHHQ2XTR6254YJFU","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"OHHQ2XTR","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":5,"sample":[{"citing_arxiv_id":"2605.06139","citing_title":"Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex","ref_index":63,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21266","citing_title":"How Much Online RL is Enough? Informative Rollouts for Offline Preference Optimization in RLVR","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08327","citing_title":"Interactive Critique-Revision Training for Reliable Structured LLM Generation","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06139","citing_title":"Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex","ref_index":62,"is_internal_anchor":true},{"citing_arxiv_id":"2604.16995","citing_title":"SPS: Steering Probability Squeezing for Better Exploration in Reinforcement Learning for Large Language Models","ref_index":40,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OHHQ2XTR6254YJFUOZYXM4TFND","json":"https://pith.science/pith/OHHQ2XTR6254YJFUOZYXM4TFND.json","graph_json":"https://pith.science/api/pith-number/OHHQ2XTR6254YJFUOZYXM4TFND/graph.json","events_json":"https://pith.science/api/pith-number/OHHQ2XTR6254YJFUOZYXM4TFND/events.json","paper":"https://pith.science/paper/OHHQ2XTR"},"agent_actions":{"view_html":"https://pith.science/pith/OHHQ2XTR6254YJFUOZYXM4TFND","download_json":"https://pith.science/pith/OHHQ2XTR6254YJFUOZYXM4TFND.json","view_paper":"https://pith.science/paper/OHHQ2XTR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2510.00977&json=true","fetch_graph":"https://pith.science/api/pith-number/OHHQ2XTR6254YJFUOZYXM4TFND/graph.json","fetch_events":"https://pith.science/api/pith-number/OHHQ2XTR6254YJFUOZYXM4TFND/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OHHQ2XTR6254YJFUOZYXM4TFND/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OHHQ2XTR6254YJFUOZYXM4TFND/action/storage_attestation","attest_author":"https://pith.science/pith/OHHQ2XTR6254YJFUOZYXM4TFND/action/author_attestation","sign_citation":"https://pith.science/pith/OHHQ2XTR6254YJFUOZYXM4TFND/action/citation_signature","submit_replication":"https://pith.science/pith/OHHQ2XTR6254YJFUOZYXM4TFND/action/replication_record"}},"created_at":"2026-05-17T23:39:17.355495+00:00","updated_at":"2026-05-17T23:39:17.355495+00:00"}