Promoting Fairness in LLMs: Detection and Mitigation of Gender Bias

Promoting Fairness in LLMs: Detection and Mitigation of Gender Bias | 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 Promoting Fairness in LLMs: Detection and Mitigation of Gender Bias Tejansh Sachdeva, Mitaali Singhal, SONIA KHETARPAUL This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6461545/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 22 Mar, 2026 Read the published version in Knowledge and Information Systems → Version 1 posted 22 You are reading this latest preprint version Abstract As large language models (LLMs) play an important role in AI applications, it is crucial to address biases, especially gender bias, to avoid stereotypes and ensure fairness in their results. A biased LLM may result in misrepresenting information, leading to social inequality, and reduced trust. To address these challenges, this research addresses the detection and mitigation of bias, with a focus on gender bias. We employed specialized metrics— Disparity Index (DI), Idea ConsistencyScore (ICS), Thematic Consistency Score (TCS), and Zero-Shot Classification—to evaluate model behavior across sensitive factors in Hindi and English prompts. These metrics were applied to analyze responses to di-verse prompts in both Hindi and English, which helps in the detection of explicit disparities in model outputs across sensitive factors such as gender. On the basis of insights gained from these evaluations, we developed two approaches to address these biases. First, we employed prompt eng-neering to refine model outputs and mitigate bias effectively. Building on these results, we further fine-tuned the model using LoRA (Low-RankAdaptation), a resource-efficient technique, to achieve substantial reductions in bias. Initial prompt engineering reduced polarized responses by 40% and improved positive portrayals by 45%. Further bias reduction was achieved through LoRA-based fine-tuning, lowering gender bias by 37%, racial bias by 27%, and age bias by 30%. These approaches show ascalable method to achieve fairness in LLMs. Bias Detection Fairness in AI Bias Mitigation Strategies Large Language Models Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 22 Mar, 2026 Read the published version in Knowledge and Information Systems → Version 1 posted Editorial decision: Revision requested 11 Jun, 2025 Reviews received at journal 09 Jun, 2025 Reviews received at journal 05 Jun, 2025 Reviews received at journal 03 Jun, 2025 Reviews received at journal 27 May, 2025 Reviews received at journal 21 May, 2025 Reviewers agreed at journal 13 May, 2025 Reviewers agreed at journal 10 May, 2025 Reviewers agreed at journal 10 May, 2025 Reviewers agreed at journal 10 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviews received at journal 08 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers invited by journal 08 May, 2025 Editor assigned by journal 18 Apr, 2025 Submission checks completed at journal 16 Apr, 2025 First submitted to journal 16 Apr, 2025 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-6461545","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":454792105,"identity":"b4336b10-2437-4ff6-b2f1-828fb423022e","order_by":0,"name":"Tejansh Sachdeva","email":"","orcid":"","institution":"Shiv Nadar Institution of Eminence, Delhi NCR","correspondingAuthor":false,"prefix":"","firstName":"Tejansh","middleName":"","lastName":"Sachdeva","suffix":""},{"id":454792106,"identity":"27927737-fe30-46bf-8cd7-1c27fa0ee0a9","order_by":1,"name":"Mitaali Singhal","email":"","orcid":"","institution":"Shiv Nadar Institution of Eminence, Delhi NCR","correspondingAuthor":false,"prefix":"","firstName":"Mitaali","middleName":"","lastName":"Singhal","suffix":""},{"id":454792107,"identity":"88c32eae-7216-415d-8316-9782e80db2c3","order_by":2,"name":"SONIA KHETARPAUL","email":"data:image/png;base64,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","orcid":"","institution":"Shiv Nadar Institution of Eminence, Delhi NCR","correspondingAuthor":true,"prefix":"","firstName":"SONIA","middleName":"","lastName":"KHETARPAUL","suffix":""}],"badges":[],"createdAt":"2025-04-16 08:53:28","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6461545/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6461545/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s10115-026-02731-3","type":"published","date":"2026-03-22T15:57:33+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":105224134,"identity":"578ad485-968b-4221-9c80-50096d1c2082","added_by":"auto","created_at":"2026-03-23 16:12:36","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":637964,"visible":true,"origin":"","legend":"","description":"","filename":"Biasness2025.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6461545/v1_covered_c8e18447-3cb5-4da7-a47f-6f7398dd4e52.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Promoting Fairness in LLMs: Detection and Mitigation of Gender Bias","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"knowledge-and-information-systems","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"kais","sideBox":"Learn more about [Knowledge and Information Systems](http://link.springer.com/journal/10115)","snPcode":"10115","submissionUrl":"https://submission.nature.com/new-submission/10115/3","title":"Knowledge and Information Systems","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Bias Detection, Fairness in AI, Bias Mitigation Strategies, Large Language Models","lastPublishedDoi":"10.21203/rs.3.rs-6461545/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6461545/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"As large language models (LLMs) play an important role in AI applications, it is crucial to address biases, especially gender bias, to avoid stereotypes and ensure fairness in their results. A biased LLM may result in misrepresenting information, leading to social inequality, and reduced trust. To address these challenges, this research addresses the detection and mitigation of bias, with a focus on gender bias. We employed specialized metrics— Disparity Index (DI), Idea ConsistencyScore (ICS), Thematic Consistency Score (TCS), and Zero-Shot Classification—to evaluate model behavior across sensitive factors in Hindi and English prompts. These metrics were applied to analyze responses to di-verse prompts in both Hindi and English, which helps in the detection of explicit disparities in model outputs across sensitive factors such as gender. On the basis of insights gained from these evaluations, we developed two approaches to address these biases. First, we employed prompt eng-neering to refine model outputs and mitigate bias effectively. Building on these results, we further fine-tuned the model using LoRA (Low-RankAdaptation), a resource-efficient technique, to achieve substantial reductions in bias. Initial prompt engineering reduced polarized responses by 40% and improved positive portrayals by 45%. Further bias reduction was achieved through LoRA-based fine-tuning, lowering gender bias by 37%, racial bias by 27%, and age bias by 30%. These approaches show ascalable method to achieve fairness in LLMs.","manuscriptTitle":"Promoting Fairness in LLMs: Detection and Mitigation of Gender Bias","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-13 10:28:19","doi":"10.21203/rs.3.rs-6461545/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-06-11T14:14:07+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-09T12:36:03+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-05T16:18:05+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-03T07:12:02+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-28T03:05:50+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-21T09:31:23+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"308711694717960474209564998773117307342","date":"2025-05-13T16:01:27+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"46156745094002069409534454435635353453","date":"2025-05-11T00:54:59+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"285814335353664163330733433042716314449","date":"2025-05-10T12:33:49+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"74826965650888343626797484490198576225","date":"2025-05-10T10:09:34+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"217855071639468339881529416386112781323","date":"2025-05-08T12:55:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"267836466813387146225532655305505290171","date":"2025-05-08T11:54:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"199997311419733849771606721570143337480","date":"2025-05-08T11:08:35+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"284562819104552639237151573001506344219","date":"2025-05-08T10:32:50+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"328130327457834127179971752897733894447","date":"2025-05-08T07:23:08+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-08T07:17:46+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"156153792015440130080470887846371753614","date":"2025-05-08T07:13:27+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"123078777160460381034475430825174772128","date":"2025-05-08T07:05:29+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-08T06:55:01+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-18T14:11:54+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-16T10:28:18+00:00","index":"","fulltext":""},{"type":"submitted","content":"Knowledge and Information Systems","date":"2025-04-16T08:49:59+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"knowledge-and-information-systems","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"kais","sideBox":"Learn more about [Knowledge and Information Systems](http://link.springer.com/journal/10115)","snPcode":"10115","submissionUrl":"https://submission.nature.com/new-submission/10115/3","title":"Knowledge and Information Systems","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"8e37b427-dd8d-4d95-92cc-4a3b23345457","owner":[],"postedDate":"May 13th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-03-23T16:09:11+00:00","versionOfRecord":{"articleIdentity":"rs-6461545","link":"https://doi.org/10.1007/s10115-026-02731-3","journal":{"identity":"knowledge-and-information-systems","isVorOnly":false,"title":"Knowledge and Information Systems"},"publishedOn":"2026-03-22 15:57:33","publishedOnDateReadable":"March 22nd, 2026"},"versionCreatedAt":"2025-05-13 10:28:19","video":"","vorDoi":"10.1007/s10115-026-02731-3","vorDoiUrl":"https://doi.org/10.1007/s10115-026-02731-3","workflowStages":[]},"version":"v1","identity":"rs-6461545","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6461545","identity":"rs-6461545","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}