Screening Feedback for Language Models with Costly Verification

Screening Feedback for Language Models with Costly Verification | 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 Article Screening Feedback for Language Models with Costly Verification Zhonglin Liu, Jussi Keppo, Murari Mandal, Hong Ming Tan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8771074/v2 This work is licensed under a CC BY 4.0 License Status: Posted Version 2 posted You are reading this latest preprint version Show more versions Abstract Language model training and alignment rely on high-quality human feedback, yet platforms must incentivize valuable contributions while limiting harmful feedback from non-experts. We study a simple screening environment in which a platform commits to a uniform incentive policy $(\rho,R,P)$---a verification rate, a reward for submitting feedback, and a penalty imposed when verified feedback is harmful---and heterogeneous users decide whether to participate. High-type users are more likely to produce helpful feedback, while low-type users are more likely to generate harmful feedback. We characterize the platform-optimal reward--penalty policy under costly verification in the robust pure-participation regimes. A key boundary condition, $\phi_H(1-\eta_H)=\phi_L(1-\eta_L)$, separates parameter regions in which incentives implement normal separation (high types participate and low types abstain) from regions exhibiting reverse screening (low types participate while high types are deterred). Verification is the primary instrument: optimal policies feature full verification at moderate costs and selective verification as verification becomes more expensive. The optimal policy is typically penalty-constrained, with rewards pinned down by participation incentives and penalties limited by enforcement and reputational costs. We further show that, under optimal verification, platform profit need not increase monotonically with population quality: profits can follow an inverted-U pattern, peaking at intermediate shares of high-type users. The mechanism is that higher population quality induces the platform to reduce verification intensity, and the resulting cost savings may be insufficient to offset foregone verification benefits. Finally, we provide an illustrative simulation using a bigram language model as a transparent calibration exercise to generate plausible magnitudes for $(\eta_H,\eta_L)$ and to visualize the model's comparative statics. Physical sciences/Engineering Physical sciences/Mathematics and computing Costly verification Incentives User feedback Quality control Platform economics Language models Full Text Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 2 posted You are reading this latest preprint version Show more versions 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-8771074","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":627669477,"identity":"590f20f8-2864-4617-9981-efa7f7c7132a","order_by":0,"name":"Zhonglin Liu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+ElEQVRIie3PMUvEMBTA8VeETk+zvtJ+iJyFckM5v0qPwLulg+MNIgUhLqLrfRTHnIG4RFw7uEjB3V0OryAuR+7OzSH/IZAHP/ICEIv9x8TD+/CFNYrfCaUASdeECfkTiQUXWfdzP0ygTQnrp1qaY4k0L46oNVg+WzdcLmfXkGum5JHDxN6rqfRvWHlelCuvCArnKPHtnlfMpJ/ffWDVY5WfarNdbKEp0cs9pJG03lgsVyPZHEXa86xDi5JG0o2Et4vp8GJZ71QJyEieOUenMk2spnMd/v7Z6816AKwvxK11OV7NhCCe9J9aBclu6Xg0fwCxWCwW2+0b1/9PiFkyYQwAAAAASUVORK5CYII=","orcid":"","institution":"The University of Hong Kong","correspondingAuthor":true,"prefix":"","firstName":"Zhonglin","middleName":"","lastName":"Liu","suffix":""},{"id":627669478,"identity":"78cd9d77-b24f-48de-88ea-42199a99f4dd","order_by":1,"name":"Jussi Keppo","email":"","orcid":"","institution":"National University of Singapore","correspondingAuthor":false,"prefix":"","firstName":"Jussi","middleName":"","lastName":"Keppo","suffix":""},{"id":627669479,"identity":"1464a988-1306-408b-8a69-71d403ba6d78","order_by":2,"name":"Murari Mandal","email":"","orcid":"","institution":"Kalinga Institute of Industrial Technology","correspondingAuthor":false,"prefix":"","firstName":"Murari","middleName":"","lastName":"Mandal","suffix":""},{"id":627669480,"identity":"0cae4757-de7d-420f-9480-6588d3a241bf","order_by":3,"name":"Hong Ming Tan","email":"","orcid":"","institution":"National University of Singapore","correspondingAuthor":false,"prefix":"","firstName":"Hong","middleName":"Ming","lastName":"Tan","suffix":""}],"badges":[],"createdAt":"2026-02-03 05:08:40","currentVersionCode":2,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8771074/v2","doiUrl":"https://doi.org/10.21203/rs.3.rs-8771074/v2","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108806354,"identity":"5f3bba33-4cdd-4cec-8ff7-896d01abf179","added_by":"auto","created_at":"2026-05-08 15:28:21","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":664868,"visible":true,"origin":"","legend":"","description":"","filename":"snarticle.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8771074/v2_covered_70a68753-b0ff-423a-93aa-f8019f6ca585.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eScreening Feedback for Language Models with Costly Verification\u003c/p\u003e","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"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":"Costly verification, Incentives, User feedback, Quality control, Platform economics, Language models","lastPublishedDoi":"10.21203/rs.3.rs-8771074/v2","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8771074/v2","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLanguage model training and alignment rely on high-quality human feedback, yet platforms must incentivize valuable contributions while limiting harmful feedback from non-experts. We study a simple screening environment in which a platform commits to a uniform incentive policy $(\\rho,R,P)$---a verification rate, a reward for submitting feedback, and a penalty imposed when verified feedback is harmful---and heterogeneous users decide whether to participate. High-type users are more likely to produce helpful feedback, while low-type users are more likely to generate harmful feedback.\u003c/p\u003e\n\u003cp\u003e\t\t\u003c/p\u003e\n\u003cp\u003e\t\tWe characterize the platform-optimal reward--penalty policy under costly verification in the robust pure-participation regimes. A key boundary condition, $\\phi_H(1-\\eta_H)=\\phi_L(1-\\eta_L)$, separates parameter regions in which incentives implement normal separation (high types participate and low types abstain) from regions exhibiting reverse screening (low types participate while high types are deterred). Verification is the primary instrument: optimal policies feature full verification at moderate costs and selective verification as verification becomes more expensive. The optimal policy is typically penalty-constrained, with rewards pinned down by participation incentives and penalties limited by enforcement and reputational costs.\u003c/p\u003e\n\u003cp\u003e\t\t\u003c/p\u003e\n\u003cp\u003e\t\tWe further show that, under optimal verification, platform profit need not increase monotonically with population quality: profits can follow an inverted-U pattern, peaking at intermediate shares of high-type users. The mechanism is that higher population quality induces the platform to reduce verification intensity, and the resulting cost savings may be insufficient to offset foregone verification benefits. Finally, we provide an illustrative simulation using a bigram language model as a transparent calibration exercise to generate plausible magnitudes for $(\\eta_H,\\eta_L)$ and to visualize the model's comparative statics.\u003c/p\u003e","manuscriptTitle":"Screening Feedback for Language Models with Costly Verification","msid":"","msnumber":"","nonDraftVersions":[{"code":2,"date":"2026-05-06 17:35:50","doi":"10.21203/rs.3.rs-8771074/v2","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}},{"code":1,"date":"2026-04-22 10:38:00","doi":"10.21203/rs.3.rs-8771074/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":"14f9d045-2360-4807-b704-01978238b19c","owner":[],"postedDate":"May 6th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":66819186,"name":"Physical sciences/Engineering"},{"id":66819187,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-04-22T10:38:00+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-06 17:35:50","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v2","identity":"rs-8771074","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8771074","identity":"rs-8771074","version":["v2"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}