GRPO++: A Controlled Differentiable Reinforcement Learning Reranking Method for Retrieval-Augmented Generation

GRPO++: A Controlled Differentiable Reinforcement Learning Reranking Method for Retrieval-Augmented Generation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article GRPO++: A Controlled Differentiable Reinforcement Learning Reranking Method for Retrieval-Augmented Generation Pengkun Zhou, Chao Gao, Xinzhe Huang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8088895/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The performance of retrieval-augmented generation (RAG) systems heavily depends on the quality of the re-ranking module, yet existing methods face challenges such as training instability, unreasonable reward design, and insufficient confidence calibration. This paper proposes GRPO++, an end-to-end re-ranking framework based on reinforcement learning, which collaboratively optimizes five core modules: Soft Normalized Discounted Cumulative Gain (Soft-nDCG) to achieve a differentiable ranking objective; Critical Reweighting for Importance (CRI) to enhance Top-K position optimization; Evidence Consistency Scoring (EGS) to ensure citation accuracy; Quartile Mean Baseline (QMB) combined with Median Absolute Deviation (MAD) for robust advantage estimation; and a KL-PID controller for adaptive trust region constraint adjustment. Experimental results on MS MARCO and BEIR multi-domain benchmarks demonstrate that GRPO++ outperforms supervised and PPO/GRPO baselines in relevance metrics such as nDCG@10, MRR@10, and MAP, while effectively reducing ECE/Brier scores and improving EGS-F1. Analysis indicates that GRPO++ significantly enhances confidence calibration and evidence consistency while maintaining high performance, providing a novel approach for reinforcement learning-based re-ranking. Retrieval-Augmented Generation Text Re-ranking Groupwise Relative Policy Optimization Reinforcement Learning Evidence Consistency Confidence Calibration Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8088895","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":629407266,"identity":"ee3431cc-b5aa-4daa-a039-4f35248c4794","order_by":0,"name":"Pengkun Zhou","email":"","orcid":"","institution":"Anhui Institute of Information Technology","correspondingAuthor":false,"prefix":"","firstName":"Pengkun","middleName":"","lastName":"Zhou","suffix":""},{"id":629407267,"identity":"4722a8bb-a7e5-4cd1-a062-781bcb901142","order_by":1,"name":"Chao Gao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA10lEQVRIiWNgGAWjYDACCQglx8Z8uAFIJxCvxZiNLbGB4QApWhIbiNbCP7v54cMvNTbpfWyMbdIfKtIY+Nu78euTuHPM2FjmWFpuG1CLxIEzOQwSZ85uwKvFQCLBTFqC7XBum3xjm8TBtgqgSC4hLenfpCX+HU5nA9ly8B9RWnLMJD+2HU6AaGnIIaxF4kZOsTFjX5oh0C/NFmeOpfEQ9Av/jPSND398s5GXb2M+eKOiJlmOv70XvxYQYOaB0CygOOIhqBwEGH9AtX4gSvkoGAWjYBSMOAAAeS9GVYP2fjMAAAAASUVORK5CYII=","orcid":"","institution":"Anhui Institute of Information Technology","correspondingAuthor":true,"prefix":"","firstName":"Chao","middleName":"","lastName":"Gao","suffix":""},{"id":629407268,"identity":"59e9ef04-a250-4dc3-8b73-1efdab647970","order_by":2,"name":"Xinzhe Huang","email":"","orcid":"","institution":"Anhui Institute of Information Technology","correspondingAuthor":false,"prefix":"","firstName":"Xinzhe","middleName":"","lastName":"Huang","suffix":""}],"badges":[],"createdAt":"2025-11-11 16:38:44","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8088895/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8088895/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108977254,"identity":"fd34f3cf-3294-4ed1-9757-2aacaf51d63e","added_by":"auto","created_at":"2026-05-11 11:31:05","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1948758,"visible":true,"origin":"","legend":"","description":"","filename":"AControlledDifferentiableReinforcementLearningRerankingMethodforRetrievalAugmentedGeneration.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8088895/v1_covered_5f522c11-5e75-4bd3-85ad-1c4bf64cb8a4.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"GRPO++: A Controlled Differentiable Reinforcement Learning Reranking Method for Retrieval-Augmented Generation","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Retrieval-Augmented Generation, Text Re-ranking, Groupwise Relative Policy Optimization, Reinforcement Learning, Evidence Consistency, Confidence Calibration","lastPublishedDoi":"10.21203/rs.3.rs-8088895/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8088895/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"The performance of retrieval-augmented generation (RAG) systems heavily depends on the quality of the re-ranking module, yet existing methods face challenges such as training instability, unreasonable reward design, and insufficient confidence calibration. This paper proposes GRPO++, an end-to-end re-ranking framework based on reinforcement learning, which collaboratively optimizes five core modules: Soft Normalized Discounted Cumulative Gain (Soft-nDCG) to achieve a differentiable ranking objective; Critical Reweighting for Importance (CRI) to enhance Top-K position optimization; Evidence Consistency Scoring (EGS) to ensure citation accuracy; Quartile Mean Baseline (QMB) combined with Median Absolute Deviation (MAD) for robust advantage estimation; and a KL-PID controller for adaptive trust region constraint adjustment. Experimental results on MS MARCO and BEIR multi-domain benchmarks demonstrate that GRPO++ outperforms supervised and PPO/GRPO baselines in relevance metrics such as nDCG@10, MRR@10, and MAP, while effectively reducing ECE/Brier scores and improving EGS-F1. Analysis indicates that GRPO++ significantly enhances confidence calibration and evidence consistency while maintaining high performance, providing a novel approach for reinforcement learning-based re-ranking.","manuscriptTitle":"GRPO++: A Controlled Differentiable Reinforcement Learning Reranking Method for Retrieval-Augmented Generation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-27 08:41:59","doi":"10.21203/rs.3.rs-8088895/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"fb530b8d-81f3-41b9-a13a-2a4fb541e49e","owner":[],"postedDate":"April 27th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-05-09T16:02:58+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-09T07:16:10+00:00","index":140,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-07T11:54:46+00:00","index":139,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-09T16:10:27+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-27 08:41:59","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8088895","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8088895","identity":"rs-8088895","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}