Expert-level validation of AI-generated medical text with scalable language models

Expert-level validation of AI-generated medical text with scalable language models | 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 Expert-level validation of AI-generated medical text with scalable language models Asad Aali, Vasiliki Bikia, Maya Varma, Nicole Chiou, Sophie Ostmeier, and 22 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7041107/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 With the growing use of language models (LMs) in clinical environments, there is an immediate need to evaluate the accuracy and safety of LM-generated medical text. Currently, such evaluation relies solely on manual physician review. However, detecting errors in LM-generated text is challenging because 1) manual review is costly and 2) expert-composed reference outputs are often unavailable in real-world settings. While the “LM-as-judge” paradigm (a LM evaluating another LM) offers scalable evaluation, even frontier LMs can miss subtle but clinically significant errors. To address these challenges, we propose MedVAL, a self-supervised framework that leverages synthetic data to train evaluator LMs to assess whether LM-generated medical outputs are factually consistent with inputs, without requiring physician labels or reference outputs. To evaluate LM performance, we introduce MedVAL-Bench, a dataset of 840 physician-annotated outputs across 6 diverse medical tasks capturing real-world challenges, including a multilingual task reviewed by bilingual physicians. Each output is reviewed following a physician-defined taxonomy of risk levels and error categories, enabling evaluation of LMs in making safety decisions for deployment. Across 10 state-of-the-art LMs spanning open-source, proprietary, and medically adapted models, MedVAL fine-tuning significantly improves (p < 0.001) alignment with physicians on both seen and unseen tasks, increasing average F1 scores from 66% to 83%, with per-sample safety classification scores up to 86%. MedVAL improves the performance of even the best-performing proprietary LM (GPT-4o) by 8%. To support a scalable, risk-aware pathway towards clinical integration, we open-source the 1) codebase, 2) MedVAL-Bench, and 3) MedVAL-4B, the best-performing open-source LM. Our research provides the first evidence of LMs approaching expert-level validation ability for medical text. Health sciences/Health care Physical sciences/Engineering/Biomedical engineering Full Text Additional Declarations There is NO Competing Interest. 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-7041107","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":481617619,"identity":"d010ab59-ad9b-4440-b884-3fbbcd865e4f","order_by":0,"name":"Asad Aali","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7ElEQVRIiWNgGAWjYJACCQYGCx5+MNMGiA+ARQhqkeCRbGAGMtNI0MJgcIBYLfz9hx/e+FEjIWN8/vzBzwUJNvl8B5gP3ubBZ8ONNGPLnmMSPGY3kpmlZySkWc48wJZsjU8Lww0GMwkeNpAWZgZp3h+HDQwO8JhJ49Mif/74N8k//yR4jPsPM//mSQBp4f+GV4vBgRwzad42CR4DhmQ2aYgWHja8Wgxv5BRby/ZJ8EjcSDaz5klIM5A8zGZsOQePFrnzxzfefPPNxp6//+Dj2zwJNgZ8x5sf3niDz/uYgJk05aNgFIyCUTAKsAAA+VtGJGumeAkAAAAASUVORK5CYII=","orcid":"https://orcid.org/0009-0008-2120-5722","institution":"Stanford University","correspondingAuthor":true,"prefix":"","firstName":"Asad","middleName":"","lastName":"Aali","suffix":""},{"id":481617620,"identity":"c8eb04ec-86cc-4f93-b3b5-84d0cc2ec2c1","order_by":1,"name":"Vasiliki Bikia","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Vasiliki","middleName":"","lastName":"Bikia","suffix":""},{"id":481617621,"identity":"d2c99805-c4f6-4fd7-9fa9-01dbfc5d81bb","order_by":2,"name":"Maya Varma","email":"","orcid":"https://orcid.org/0000-0003-0693-7753","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Maya","middleName":"","lastName":"Varma","suffix":""},{"id":481617622,"identity":"7281ed66-e9a8-4e57-a8fe-cd8b4667f157","order_by":3,"name":"Nicole Chiou","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Nicole","middleName":"","lastName":"Chiou","suffix":""},{"id":481617623,"identity":"7857492e-0c27-4840-9f1f-38e7e6e9d787","order_by":4,"name":"Sophie Ostmeier","email":"","orcid":"","institution":"Stanford","correspondingAuthor":false,"prefix":"","firstName":"Sophie","middleName":"","lastName":"Ostmeier","suffix":""},{"id":481617624,"identity":"e0f67ba7-08c8-434e-98f4-726e5a157e21","order_by":5,"name":"Arnav Singhvi","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Arnav","middleName":"","lastName":"Singhvi","suffix":""},{"id":481617625,"identity":"dd32356e-a785-4711-a883-545ac386413a","order_by":6,"name":"Magdalini Paschali","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Magdalini","middleName":"","lastName":"Paschali","suffix":""},{"id":481617626,"identity":"b72eeedd-5744-49e6-8bef-e9bd58397bed","order_by":7,"name":"Ashwin Kumar","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Ashwin","middleName":"","lastName":"Kumar","suffix":""},{"id":481617627,"identity":"14f03767-ded9-4ea3-99f1-2845b3316e12","order_by":8,"name":"Andrew Johnston","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Andrew","middleName":"","lastName":"Johnston","suffix":""},{"id":481617628,"identity":"765c832b-f399-457d-bccd-c2c9b83c2f12","order_by":9,"name":"Karimar Amador-Martinez","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Karimar","middleName":"","lastName":"Amador-Martinez","suffix":""},{"id":481617629,"identity":"42b32055-d366-423d-89f4-9b3300314813","order_by":10,"name":"Eduardo Guerrero","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Eduardo","middleName":"","lastName":"Guerrero","suffix":""},{"id":481617630,"identity":"0b48c1d8-b4cf-4e64-8e3d-21d76da2a2e9","order_by":11,"name":"Paola Rivera","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Paola","middleName":"","lastName":"Rivera","suffix":""},{"id":481617631,"identity":"047bbf32-b43c-4512-a6c2-5bb4be6e737b","order_by":12,"name":"Sergios Gatidis","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Sergios","middleName":"","lastName":"Gatidis","suffix":""},{"id":481617632,"identity":"4b84219d-1bed-4b4c-9a09-5b39943badd6","order_by":13,"name":"Christian Bluethgen","email":"","orcid":"https://orcid.org/0000-0001-7321-5676","institution":"University Hospital Zurich","correspondingAuthor":false,"prefix":"","firstName":"Christian","middleName":"","lastName":"Bluethgen","suffix":""},{"id":481617633,"identity":"01d02276-a9c4-491c-8383-8ade80b81053","order_by":14,"name":"Eduardo Pontes Reis","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Eduardo","middleName":"Pontes","lastName":"Reis","suffix":""},{"id":481617634,"identity":"551231f4-21dc-4328-9b6e-eb812c71c4dc","order_by":15,"name":"Eddy Rilland","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Eddy","middleName":"","lastName":"Rilland","suffix":""},{"id":481617635,"identity":"e7d6dd20-67fc-4663-8230-5e5a737f3537","order_by":16,"name":"Poonam Hosamani","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Poonam","middleName":"","lastName":"Hosamani","suffix":""},{"id":481617636,"identity":"c64ce18e-6233-4594-af3f-69ea799c703e","order_by":17,"name":"Kevin Keet","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Kevin","middleName":"","lastName":"Keet","suffix":""},{"id":481617637,"identity":"d64d0ea6-39d2-4f4c-9e42-9e8c3072010c","order_by":18,"name":"Minjoung Go","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Minjoung","middleName":"","lastName":"Go","suffix":""},{"id":481617638,"identity":"78152f3b-6f1b-49a3-9b06-818d0c9aedf6","order_by":19,"name":"Evelyn Bin Ling","email":"","orcid":"","institution":"Department of Internal Medicine, Stanford University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Evelyn","middleName":"Bin","lastName":"Ling","suffix":""},{"id":481617639,"identity":"42f8b168-31fe-4a01-afa3-676f33d9486b","order_by":20,"name":"David Larson","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"","lastName":"Larson","suffix":""},{"id":481617640,"identity":"4e04fbd1-4e43-4840-af67-ce8b80e7442b","order_by":21,"name":"Curtis Langlotz","email":"","orcid":"https://orcid.org/0000-0002-8972-8051","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Curtis","middleName":"","lastName":"Langlotz","suffix":""},{"id":481617645,"identity":"c54b5551-81f1-435d-be26-4a667749852e","order_by":22,"name":"Roxana Daneshjou","email":"","orcid":"https://orcid.org/0000-0001-7988-9356","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Roxana","middleName":"","lastName":"Daneshjou","suffix":""},{"id":481617646,"identity":"be772b33-e19d-4a2b-9dee-e0afe38b5ba0","order_by":23,"name":"Jason Hom","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Jason","middleName":"","lastName":"Hom","suffix":""},{"id":481617647,"identity":"91c98a02-e083-4cd6-b578-84eaf7a9dbf7","order_by":24,"name":"Sanmi Koyejo","email":"","orcid":"https://orcid.org/0000-0002-4023-419X","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Sanmi","middleName":"","lastName":"Koyejo","suffix":""},{"id":481617648,"identity":"4a727534-1eba-486f-9f02-a11db2072857","order_by":25,"name":"Emily Alsentzer","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Emily","middleName":"","lastName":"Alsentzer","suffix":""},{"id":481617649,"identity":"794050ae-270a-4b58-a1de-dcd97c602b98","order_by":26,"name":"Akshay Chaudhari","email":"","orcid":"https://orcid.org/0000-0002-3667-6796","institution":"Stanford","correspondingAuthor":false,"prefix":"","firstName":"Akshay","middleName":"","lastName":"Chaudhari","suffix":""}],"badges":[],"createdAt":"2025-07-03 19:50:14","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7041107/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7041107/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":93243656,"identity":"831b32e5-aa17-47e6-ad32-cb7cf856b209","added_by":"auto","created_at":"2025-10-10 14:59:47","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7062982,"visible":true,"origin":"","legend":"Article File","description":"","filename":"MedVAL.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7041107/v1_covered_6a8c6061-6d90-4a14-94a8-15cb40e84c03.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Expert-level validation of AI-generated medical text with scalable language models","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":true,"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":"","lastPublishedDoi":"10.21203/rs.3.rs-7041107/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7041107/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"With the growing use of language models (LMs) in clinical environments, there is an immediate need to evaluate the accuracy and safety of LM-generated medical text. Currently, such evaluation relies solely on manual physician review. However, detecting errors in LM-generated text is challenging because 1) manual review is costly and 2) expert-composed reference outputs are often unavailable in real-world settings. While the “LM-as-judge” paradigm (a LM evaluating another LM) offers scalable evaluation, even frontier LMs can miss subtle but clinically significant errors. To address these challenges, we propose MedVAL, a self-supervised framework that leverages synthetic data to train evaluator LMs to assess whether LM-generated medical outputs are factually consistent with inputs, without requiring physician labels or reference outputs. To evaluate LM performance, we introduce MedVAL-Bench, a dataset of 840 physician-annotated outputs across 6 diverse medical tasks capturing real-world challenges, including a multilingual task reviewed by bilingual physicians. Each output is reviewed following a physician-defined taxonomy of risk levels and error categories, enabling evaluation of LMs in making safety decisions for deployment. Across 10 state-of-the-art LMs spanning open-source, proprietary, and medically adapted models, MedVAL fine-tuning significantly improves (p \u003c 0.001) alignment with physicians on both seen and unseen tasks, increasing average F1 scores from 66% to 83%, with per-sample safety classification scores up to 86%. MedVAL improves the performance of even the best-performing proprietary LM (GPT-4o) by 8%. To support a scalable, risk-aware pathway towards clinical integration, we open-source the 1) codebase, 2) MedVAL-Bench, and 3) MedVAL-4B, the best-performing open-source LM. Our research provides the first evidence of LMs approaching expert-level validation ability for medical text.","manuscriptTitle":"Expert-level validation of AI-generated medical text with scalable language models","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-08 14:23:35","doi":"10.21203/rs.3.rs-7041107/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":"5cb70c3a-77ce-4c7f-82f1-9a778ded3b52","owner":[],"postedDate":"July 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":51133147,"name":"Health sciences/Health care"},{"id":51133148,"name":"Physical sciences/Engineering/Biomedical engineering"}],"tags":[],"updatedAt":"2025-10-23T12:23:25+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-08 14:23:35","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7041107","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7041107","identity":"rs-7041107","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}