{"paper":{"title":"F-GRPO: Factorized Group-Relative Policy Optimization for Unified Candidate Generation and Ranking","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"F-GRPO lets one LLM jointly generate candidates and rank them by factorizing policy optimization into separate phases with distinct advantages.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bowen Jin, Gagan Mundada, Jiawei Han, Jingbo Shang, Julian McAuley, Junda Wu, Ritwik Sinha, Rohan Surana, Sizhe Zhou, Tong Yu, Xintong Li, Yizhu Jiao","submitted_at":"2026-05-13T04:52:33Z","abstract_excerpt":"Traditional retrieval pipelines optimize utility through stages of candidate retrieval and reranking, where ranking operates over a predefined candidate set. Large Language Models (LLMs) broaden this into a generative process: given a candidate pool, an LLM can generate a subset and order it within a single autoregressive pass. However, this flexibility introduces a new optimization challenge: the model must search a combinatorial output space while receiving utility feedback only after the full ranked list is generated. Because this feedback is defined over the completed sequence, it cannot d"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"F-GRPO improves top-ranked performance over GRPO and decoupled baselines, outperforms supervised alternatives, and remains competitive with strong zero-shot rerankers, with no architectural changes at inference time.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The phase-specific credit assignment problem can be resolved by applying separate group-relative advantages to generation and ranking inside a two-phase sequence-level objective while sharing a single LLM backbone.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"F-GRPO factorizes group-relative policy optimization into generation and ranking phases within one autoregressive sequence, using order-invariant coverage and position-aware utility rewards to improve top-ranked performance on recommendation and multi-hop QA tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"F-GRPO lets one LLM jointly generate candidates and rank them by factorizing policy optimization into separate phases with distinct advantages.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0dbd9a6a253db09894895d4c13d27b6bc4e53f14e637762b3f01342e8dd2c1a6"},"source":{"id":"2605.12995","kind":"arxiv","version":1},"verdict":{"id":"d459ee2b-4a28-4298-8c4a-93e7cced2add","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:48:31.789661Z","strongest_claim":"F-GRPO improves top-ranked performance over GRPO and decoupled baselines, outperforms supervised alternatives, and remains competitive with strong zero-shot rerankers, with no architectural changes at inference time.","one_line_summary":"F-GRPO factorizes group-relative policy optimization into generation and ranking phases within one autoregressive sequence, using order-invariant coverage and position-aware utility rewards to improve top-ranked performance on recommendation and multi-hop QA tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The phase-specific credit assignment problem can be resolved by applying separate group-relative advantages to generation and ranking inside a two-phase sequence-level objective while sharing a single LLM backbone.","pith_extraction_headline":"F-GRPO lets one LLM jointly generate candidates and rank them by factorizing policy optimization into separate phases with distinct advantages."},"references":{"count":76,"sample":[{"doi":"","year":2011,"title":"Ellis, Brian Whitman, and Paul Lamere","work_id":"3fbb6e4b-5eaa-4672-b118-27365be53cc8","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Autoregressive search engines: Generating substrings as document identifiers","work_id":"c548f215-b75f-4910-848d-3fe1c6887e1d","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1145/3539597.3570412","year":2023,"title":"Generative slate recommendation with reinforcement learning","work_id":"77c5c94e-635e-4641-b5d7-67678c66e18e","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Chang, Claire Cardie, Kianté Brantley, and Thorsten Joachim","work_id":"6a5e62e2-6e43-4821-b297-5869267b9c03","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2022.naacl-main.194","year":2022,"title":"URL https://doi.org/10.18653/v1/2022.naa cl-main.194","work_id":"57971df4-d73b-49da-9f92-43d0c13f54c9","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":76,"snapshot_sha256":"aff0f130a45adbc22583b2d53d98501c106a20ed972a6a0cc81c55469c4a8199","internal_anchors":7},"formal_canon":{"evidence_count":2,"snapshot_sha256":"53636b18e9bedcc33a89fcdfd91754292b351dd370df35424c9ff5fcdf84d34f"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}