{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:FUFM5XAVNF56WGVZEI6EYDE4BB","short_pith_number":"pith:FUFM5XAV","schema_version":"1.0","canonical_sha256":"2d0acedc15697beb1ab9223c4c0c9c08447de41a62d447b60ae5f67d938c72eb","source":{"kind":"arxiv","id":"2502.18548","version":3},"attestation_state":"computed","paper":{"title":"What is the Alignment Objective of GRPO?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Milan Vojnovic, Se-Young Yun","submitted_at":"2025-02-25T15:56:56Z","abstract_excerpt":"In this note, we examine the aggregation of preferences achieved by the Group Policy Optimisation (GRPO) algorithm, a reinforcement learning method used to train advanced artificial intelligence models such as DeepSeek-R1-Zero and DeepSeekMath. The GRPO algorithm trains a policy using a reward preference model, which is computed by sampling a set of outputs for a given context, observing the corresponding rewards, and applying shift-and-scale normalisation to these reward values. Additionally, it incorporates a penalty function to discourage deviations from a reference policy.\n  We present a f"},"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":"2502.18548","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-25T15:56:56Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"e8a4c6cfbf811e6928fedb9f0888536150863d2acc2321c54f00a6628627c16d","abstract_canon_sha256":"57b436cc60cfb230ff9d15b5cae1beb6c413825983b46054b2899458abf1274e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:30:31.490851Z","signature_b64":"i1XKoY3x6/VCIhxWT8X1ZjjFUDMlESAPfSifdiwPhogpVM7OJ4QmLmUX7AKTZlnKKtE+Fih5j2OJaEUcAGarCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2d0acedc15697beb1ab9223c4c0c9c08447de41a62d447b60ae5f67d938c72eb","last_reissued_at":"2026-07-05T10:30:31.489784Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:30:31.489784Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"What is the Alignment Objective of GRPO?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Milan Vojnovic, Se-Young Yun","submitted_at":"2025-02-25T15:56:56Z","abstract_excerpt":"In this note, we examine the aggregation of preferences achieved by the Group Policy Optimisation (GRPO) algorithm, a reinforcement learning method used to train advanced artificial intelligence models such as DeepSeek-R1-Zero and DeepSeekMath. The GRPO algorithm trains a policy using a reward preference model, which is computed by sampling a set of outputs for a given context, observing the corresponding rewards, and applying shift-and-scale normalisation to these reward values. Additionally, it incorporates a penalty function to discourage deviations from a reference policy.\n  We present a f"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.18548","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":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2502.18548/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":"2502.18548","created_at":"2026-07-05T10:30:31.489930+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.18548v3","created_at":"2026-07-05T10:30:31.489930+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.18548","created_at":"2026-07-05T10:30:31.489930+00:00"},{"alias_kind":"pith_short_12","alias_value":"FUFM5XAVNF56","created_at":"2026-07-05T10:30:31.489930+00:00"},{"alias_kind":"pith_short_16","alias_value":"FUFM5XAVNF56WGVZ","created_at":"2026-07-05T10:30:31.489930+00:00"},{"alias_kind":"pith_short_8","alias_value":"FUFM5XAV","created_at":"2026-07-05T10:30:31.489930+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.30789","citing_title":"Predictable GRPO: A Closed-Form Model of Training Dynamics","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30789","citing_title":"Predictable GRPO: A Closed-Form Model of Training Dynamics","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06139","citing_title":"Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2602.05946","citing_title":"f-GRPO and Beyond: Divergence-Based Reinforcement Learning Algorithms for General LLM Alignment","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06139","citing_title":"Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2604.20755","citing_title":"V-tableR1: Process-Supervised Multimodal Table Reasoning with Critic-Guided Policy Optimization","ref_index":30,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/FUFM5XAVNF56WGVZEI6EYDE4BB","json":"https://pith.science/pith/FUFM5XAVNF56WGVZEI6EYDE4BB.json","graph_json":"https://pith.science/api/pith-number/FUFM5XAVNF56WGVZEI6EYDE4BB/graph.json","events_json":"https://pith.science/api/pith-number/FUFM5XAVNF56WGVZEI6EYDE4BB/events.json","paper":"https://pith.science/paper/FUFM5XAV"},"agent_actions":{"view_html":"https://pith.science/pith/FUFM5XAVNF56WGVZEI6EYDE4BB","download_json":"https://pith.science/pith/FUFM5XAVNF56WGVZEI6EYDE4BB.json","view_paper":"https://pith.science/paper/FUFM5XAV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.18548&json=true","fetch_graph":"https://pith.science/api/pith-number/FUFM5XAVNF56WGVZEI6EYDE4BB/graph.json","fetch_events":"https://pith.science/api/pith-number/FUFM5XAVNF56WGVZEI6EYDE4BB/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FUFM5XAVNF56WGVZEI6EYDE4BB/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FUFM5XAVNF56WGVZEI6EYDE4BB/action/storage_attestation","attest_author":"https://pith.science/pith/FUFM5XAVNF56WGVZEI6EYDE4BB/action/author_attestation","sign_citation":"https://pith.science/pith/FUFM5XAVNF56WGVZEI6EYDE4BB/action/citation_signature","submit_replication":"https://pith.science/pith/FUFM5XAVNF56WGVZEI6EYDE4BB/action/replication_record"}},"created_at":"2026-07-05T10:30:31.489930+00:00","updated_at":"2026-07-05T10:30:31.489930+00:00"}