EPMORE: Explainable Process Mixture-of-Experts

EPMORE: Explainable Process Mixture-of-Experts | 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 EPMORE: Explainable Process Mixture-of-Experts Wei Sheng, Chengzhu Xiao, Lunhao Ao, Junyan Long, Ye Yu, Yangguang Jia, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8374807/v4 This work is licensed under a CC BY 4.0 License Status: Posted Version 4 posted You are reading this latest preprint version Show more versions Abstract Large language models (LLMs) are primarily built on the Transformer architecture, in which all hidden layers share a fixed-dimensional representation spaces. This homogeneity constrains representational capacity, impedes interpretability, and induces computational redundancy. We propose EPMORE (Explainable Process Mixture-of-Experts), a novel architecture that models inference as a process of dimensional elevation, Ensure the entire inference/training process, intermediate states observable and explainable, and ensure the whole process is traceable end-to-end. EPMORE decomposes the entire inference/training process into a hierarchical sequence of representation spaces — from a semantic space (128 dimensions), to one or more logical spaces (512 dimensions each), and finally to fact-expert representation spaces (1024 dimensions) — allowing deeper network stages to encode progressively richer and more abstract features. A core component, Middle Output Reuse (MOR), enables each layer to produce interpretable intermediate predictions. Theoretically, forward propagation can be interpreted as representation-space expansion, while backward propagation corresponds to a dimensional contraction process. Experiments show that, compared with dense and conventional mixture-of-experts (MoE, Deepseek) baselines, EPMORE improves interpretability, activation sparsity, parameter independence, and inference performance while reducing computational cost. These findings suggest that hierarchical dimensional elevation is a promising alternative to standard Transformer design. Artificial Intelligence and Machine Learning Explainable Process Low Computational Complexity Weight Independence Simple Positional Embedding Hierarchical tokenizer and BPE MOR MOE Full Text Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 4 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-8374807","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":590539793,"identity":"575414a1-4e38-49bb-9cdf-99066a13c94d","order_by":0,"name":"Wei Sheng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7klEQVRIiWNgGAWjYDACCQjFww8mGSwSiNci2QDWIgHUwkycFgaDA8RqkZ/dfOzh1zYbGeNrh589eFMhkWfe3n+A4ecO3FoM7hxLN5ZtS+Mxu51mbjjnjESxzJnDDIy9Z/Bokcgxk5bcdhioJcFMmrdNInGGRDIDM2MbHofNyP8G1PKfx3h2+jdp3n9EaGG4kcMm+XHbAR4DaaB1vA1EaDG4kWYmzfgvmUfidk6Z5JxjEsUSPIcNDvbidVjyM8kfZ+zs+Wenb5N4U2OTJ8He+PDBT3wOAwJmHnSRA/g1MDAw/iCkYhSMglEwCkY2AACplEvhMf0CKQAAAABJRU5ErkJggg==","orcid":"","institution":"Mandelbrot Technology Co., LTD","correspondingAuthor":true,"prefix":"","firstName":"Wei","middleName":"","lastName":"Sheng","suffix":""},{"id":590539794,"identity":"0abfd7f1-6212-4221-9dca-c1c427b285c0","order_by":1,"name":"Chengzhu Xiao","email":"","orcid":"","institution":"Mandelbrot Technology Co., LTD","correspondingAuthor":false,"prefix":"","firstName":"Chengzhu","middleName":"","lastName":"Xiao","suffix":""},{"id":590539795,"identity":"fe654f0a-ec34-4a75-9d7b-6fa3d8f7c5d4","order_by":2,"name":"Lunhao Ao","email":"","orcid":"","institution":"Mandelbrot Technology Co., LTD","correspondingAuthor":false,"prefix":"","firstName":"Lunhao","middleName":"","lastName":"Ao","suffix":""},{"id":590539796,"identity":"1e37e7a7-2567-45d9-8596-8786eb2ec8f2","order_by":3,"name":"Junyan Long","email":"","orcid":"","institution":"Mandelbrot Technology Co., LTD","correspondingAuthor":false,"prefix":"","firstName":"Junyan","middleName":"","lastName":"Long","suffix":""},{"id":590539797,"identity":"711541e4-648c-4c45-bcf0-82e0b3298aee","order_by":4,"name":"Ye Yu","email":"","orcid":"","institution":"Mandelbrot Technology Co., LTD","correspondingAuthor":false,"prefix":"","firstName":"Ye","middleName":"","lastName":"Yu","suffix":""},{"id":590539798,"identity":"481c4c7b-c665-478a-bc13-c12547bdfd67","order_by":5,"name":"Yangguang Jia","email":"","orcid":"","institution":"Mandelbrot Technology Co., LTD","correspondingAuthor":false,"prefix":"","firstName":"Yangguang","middleName":"","lastName":"Jia","suffix":""},{"id":590539799,"identity":"ec8c5c95-d688-435d-9e69-a63608151099","order_by":6,"name":"Qihua Zhang","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Qihua","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2025-12-16 10:12:47","currentVersionCode":4,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8374807/v4","doiUrl":"https://doi.org/10.21203/rs.3.rs-8374807/v4","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105564492,"identity":"f8a95f66-f8c8-46bf-afaf-5715d0972fd6","added_by":"auto","created_at":"2026-03-27 12:49:46","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3339839,"visible":true,"origin":"","legend":"","description":"","filename":"main.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8374807/v4_covered_eae062d4-49f2-4188-96e4-0f4761dea7fb.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"EPMORE: Explainable Process Mixture-of-Experts","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Mandelbrot Technology Co., LTD","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":"Explainable Process, Low Computational Complexity, Weight Independence, Simple Positional Embedding, Hierarchical tokenizer and BPE, MOR, MOE","lastPublishedDoi":"10.21203/rs.3.rs-8374807/v4","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8374807/v4","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLarge language models (LLMs) are primarily built on the Transformer architecture, in which all hidden layers share a fixed-dimensional representation spaces. This homogeneity constrains representational capacity, impedes interpretability, and induces computational redundancy. We propose EPMORE (Explainable Process Mixture-of-Experts), a novel architecture that models inference as a process of dimensional elevation, Ensure the entire inference/training process, intermediate states observable and explainable, and ensure the whole process is traceable end-to-end. EPMORE decomposes the entire inference/training process into a hierarchical sequence of representation spaces — from a semantic space (128 dimensions), to one or more logical spaces (512 dimensions each), and finally to fact-expert representation spaces (1024 dimensions) — allowing deeper network stages to encode progressively richer and more abstract features. A core component, Middle Output Reuse (MOR), enables each layer to produce interpretable intermediate predictions. Theoretically, forward propagation can be interpreted as representation-space expansion, while backward propagation corresponds to a dimensional contraction process. Experiments show that, compared with dense and conventional mixture-of-experts (MoE, Deepseek) baselines, EPMORE improves interpretability, activation sparsity, parameter independence, and inference performance while reducing computational cost. These findings suggest that hierarchical dimensional elevation is a promising alternative to standard Transformer design.\u003c/p\u003e","manuscriptTitle":"EPMORE: Explainable Process Mixture-of-Experts","msid":"","msnumber":"","nonDraftVersions":[{"code":4,"date":"2026-03-24 18:49:44","doi":"10.21203/rs.3.rs-8374807/v4","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":3,"date":"2026-02-12 23:21:52","doi":"10.21203/rs.3.rs-8374807/v3","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":2,"date":"2026-02-05 19:02:46","doi":"10.21203/rs.3.rs-8374807/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":"2025-12-17 03:08:16","doi":"10.21203/rs.3.rs-8374807/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":"bb719e71-a9c5-4087-a7c9-d76b0b771991","owner":[],"postedDate":"March 24th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":62842642,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-12-17T03:08:16+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-24 18:49:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v4","identity":"rs-8374807","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8374807","identity":"rs-8374807","version":["v4"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}