{"paper_id":"d5df0b7e-7482-46ad-9f4c-9119080b3cd4","body_text":"Fine-grained Debiasing for Large Language Modelsvia Bias Intensity and Probability Decoupling | 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 Fine-grained Debiasing for Large Language Modelsvia Bias Intensity and Probability Decoupling Zhuge Yan, Xiaolong Gong, Wangchao Wu, Zhike Han This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8984881/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Large Language Models (LLMs) have demonstrated remarkable capabilities butoften inherit and even amplify social biases present in their training data. Existingdebiasing approaches—particularly those based on human preference alignment,such as Reinforcement Learning from Human Feedback (RLHF) and DirectPreference Optimization (DPO)—typically treat bias as a binary attribute, overlooking the nuanced differences in bias intensity. Moreover, optimizing solely forthe probability gap between preferred (less biased) and rejected (more biased)responses can lead to an undesirable phenomenon where the probabilities of bothbiased and neutral responses increase simultaneously.To address these limitations, we propose a novel fine-grained debiasing frameworkfor LLMs featuring two key innovations. First, we introduce a method to quantifybias intensity using a multi-model evaluation committee and integrate this finegrained signal into the DPO objective, resulting in Bias-Intensity Weighted DPO(BIW-DPO). This enables the model to apply differentiated penalties based onthe severity of bias. Second, we propose a Probability Decoupling Regularization(PDR) term that dynamically suppresses the probabilities of both preferred andrejected responses according to the perceived bias level, effectively preventingthe coupled escalation of biased outputs.Extensive experiments on both English and Chinese bias benchmarks (BBQ,CBBQ, GenderAlign) demonstrate that our integrated approach, DPO-FGD,achieves substantial bias reduction compared to standard DPO while mitigating performance degradation on general capability benchmarks (MMLU,GSM8K, MT-Bench). Our analysis further confirms the effectiveness of finegrained bias intensity modeling and highlights the critical role of decouplingresponse probabilities in robust debiasing. Large Language Models (LLMs) Social Bias AI Ethics Human Preference Alignment Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 10 Apr, 2026 Reviewers invited by journal 31 Mar, 2026 Editor assigned by journal 04 Mar, 2026 Submission checks completed at journal 04 Mar, 2026 First submitted to journal 27 Feb, 2026 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-8984881\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":false,\"archivedVersions\":[],\"articleType\":\"Research Article\",\"associatedPublications\":[],\"authors\":[{\"id\":616899418,\"identity\":\"4a704900-e7e4-4568-831e-a1f2a842b263\",\"order_by\":0,\"name\":\"Zhuge Yan\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Zhejiang Technical Institute of Economics\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Zhuge\",\"middleName\":\"\",\"lastName\":\"Yan\",\"suffix\":\"\"},{\"id\":616899419,\"identity\":\"9dbae885-1721-49d3-8422-2eeafa399088\",\"order_by\":1,\"name\":\"Xiaolong Gong\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Zhejiang University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Xiaolong\",\"middleName\":\"\",\"lastName\":\"Gong\",\"suffix\":\"\"},{\"id\":616899420,\"identity\":\"39a1e9f0-160a-44c3-8628-bb8562fc4f8b\",\"order_by\":2,\"name\":\"Wangchao Wu\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Hangzhou City University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Wangchao\",\"middleName\":\"\",\"lastName\":\"Wu\",\"suffix\":\"\"},{\"id\":616899421,\"identity\":\"d4edb75a-073e-4ddd-ad1e-67cd50f12f09\",\"order_by\":3,\"name\":\"Zhike Han\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAvklEQVRIiWNgGAWjYBACAyBmbKg4AOHxEK/lDMlaGttI0WLOfvbwx5nz7tjzz0hgfPC2jUHenJAWy568BMON254xS9xIYDac28ZguLOBkMMO5BgkPtx2mM1AIoFNmreNIcHgACEt598YHHw45zAPUAv7b+K03MgxbNzYcFgCZAszkVreGDPOOHbYQOLMw2bJOeckDDcQdliO8ceemsP2/O3JBz+8KbORJ2gLEmBsABISxKsfBaNgFIyCUYAbAAAxIUIk3OLHvAAAAABJRU5ErkJggg==\",\"orcid\":\"\",\"institution\":\"Hangzhou City University\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Zhike\",\"middleName\":\"\",\"lastName\":\"Han\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2026-02-27 07:40:47\",\"currentVersionCode\":1,\"declarations\":\"\",\"doi\":\"10.21203/rs.3.rs-8984881/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-8984881/v1\",\"draftVersion\":[],\"editorialEvents\":[],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":106402969,\"identity\":\"149d5160-99f4-4472-a850-1fc65ee4ae80\",\"added_by\":\"auto\",\"created_at\":\"2026-04-08 09:13:15\",\"extension\":\"pdf\",\"order_by\":1,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":3938482,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"file.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8984881/v1_covered_1b86e069-4369-4623-8021-603a6fd6dbe0.pdf\"}],\"financialInterests\":\"No competing interests reported.\",\"formattedTitle\":\"Fine-grained Debiasing for Large Language Modelsvia Bias Intensity and Probability Decoupling\",\"fulltext\":[],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":false,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":false,\"hideJournal\":false,\"highlight\":\"\",\"institution\":\"\",\"isAcceptedByJournal\":false,\"isAuthorSuppliedPdf\":true,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":true,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"neural-processing-letters\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"nepl\",\"sideBox\":\"Learn more about [Neural Processing Letters](http://link.springer.com/journal/11063)\",\"snPcode\":\"11063\",\"submissionUrl\":\"https://submission.nature.com/new-submission/11063/3\",\"title\":\"Neural Processing Letters\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"em\",\"reportingPortfolio\":\"Springer Hybrid\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":false},\"keywords\":\"Large Language Models (LLMs), Social Bias, AI Ethics, Human Preference Alignment\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-8984881/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-8984881/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"Large Language Models (LLMs) have demonstrated remarkable capabilities butoften inherit and even amplify social biases present in their training data. Existingdebiasing approaches—particularly those based on human preference alignment,such as Reinforcement Learning from Human Feedback (RLHF) and DirectPreference Optimization (DPO)—typically treat bias as a binary attribute, overlooking the nuanced differences in bias intensity. Moreover, optimizing solely forthe probability gap between preferred (less biased) and rejected (more biased)responses can lead to an undesirable phenomenon where the probabilities of bothbiased and neutral responses increase simultaneously.To address these limitations, we propose a novel fine-grained debiasing frameworkfor LLMs featuring two key innovations. First, we introduce a method to quantifybias intensity using a multi-model evaluation committee and integrate this finegrained signal into the DPO objective, resulting in Bias-Intensity Weighted DPO(BIW-DPO). This enables the model to apply differentiated penalties based onthe severity of bias. Second, we propose a Probability Decoupling Regularization(PDR) term that dynamically suppresses the probabilities of both preferred andrejected responses according to the perceived bias level, effectively preventingthe coupled escalation of biased outputs.Extensive experiments on both English and Chinese bias benchmarks (BBQ,CBBQ, GenderAlign) demonstrate that our integrated approach, DPO-FGD,achieves substantial bias reduction compared to standard DPO while mitigating performance degradation on general capability benchmarks (MMLU,GSM8K, MT-Bench). Our analysis further confirms the effectiveness of finegrained bias intensity modeling and highlights the critical role of decouplingresponse probabilities in robust debiasing.\",\"manuscriptTitle\":\"Fine-grained Debiasing for Large Language Modelsvia Bias Intensity and Probability Decoupling\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2026-04-06 13:32:22\",\"doi\":\"10.21203/rs.3.rs-8984881/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0},{\"type\":\"reviewerAgreed\",\"content\":\"242202477425604522695729201564537751235\",\"date\":\"2026-04-10T12:58:16+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewersInvited\",\"content\":\"\",\"date\":\"2026-04-01T01:38:40+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"editorAssigned\",\"content\":\"\",\"date\":\"2026-03-04T07:15:29+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"checksComplete\",\"content\":\"\",\"date\":\"2026-03-04T07:13:19+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"submitted\",\"content\":\"Neural Processing Letters\",\"date\":\"2026-02-27T07:30:47+00:00\",\"index\":\"\",\"fulltext\":\"\"}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"neural-processing-letters\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"nepl\",\"sideBox\":\"Learn more about [Neural Processing Letters](http://link.springer.com/journal/11063)\",\"snPcode\":\"11063\",\"submissionUrl\":\"https://submission.nature.com/new-submission/11063/3\",\"title\":\"Neural Processing Letters\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"em\",\"reportingPortfolio\":\"Springer Hybrid\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":false}}],\"origin\":\"\",\"ownerIdentity\":\"0911d15a-e044-49e2-a9b5-43ae0a05a7ee\",\"owner\":[],\"postedDate\":\"April 6th, 2026\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"under-review\",\"subjectAreas\":[],\"tags\":[],\"updatedAt\":\"2026-04-06T13:32:22+00:00\",\"versionOfRecord\":[],\"versionCreatedAt\":\"2026-04-06 13:32:22\",\"video\":\"\",\"vorDoi\":\"\",\"vorDoiUrl\":\"\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-8984881\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-8984881\",\"identity\":\"rs-8984881\",\"version\":[\"v1\"]},\"buildId\":\"XKTyCvWXoU3ODBz1xrDgd\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}