Efficient Prompt Compression on Edge Devices

Efficient Prompt Compression on Edge Devices | 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 Efficient Prompt Compression on Edge Devices Samyuktha D, Chandravel Saravanan, Anusha Jayasimhan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7976248/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 Large Language Models (LLMs) like GPT-4, BERT, and DeBERTa are widely used for tasks such as question answering, text summarization, and reasoning in many domains [1]–[3]. However, running these models on small or low-power devices such as mobile phones and IoT systems is challenging because they require large amounts of memory, processing power, and time when handling long text inputs called prompts. To solve this problem, researchers have developed several prompt compression methods that shorten prompts while preserving their key meaning. These include summarization-based [4], embedding-based [5], and graph-based reasoning techniques [6]. Recent methods such as LLMLingua-2 [15] and Prompt Compression with Context-Aware Sentence Encoding [20] have further improved compression quality while maintaining reasoning consistency and efficiency. Building upon these works, this paper proposes an Efficient Prompt Compression on Edge Devices framework that integrates embedding retrieval, causal-temporal reasoning, and coherence validation into a single, lightweight process. The framework produces interpretable reasoning graphs that retain only the most important information, enabling faster and more efficient processing. Experiments conducted on the CQR dataset demonstrate that the proposed model achieves high reasoning accuracy while significantly reducing computational cost, making it suitable for real-time deployment on edge devices. BLEU, and F1, making it suitable for real-world deployment on low-power devices. Artificial Intelligence and Machine Learning Large language models Prompt compression edge devices Full Text Additional Declarations The authors declare no competing interests. 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-7976248","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":536684534,"identity":"50c74309-68fa-488c-9369-be24b9ac0355","order_by":0,"name":"Samyuktha D","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Samyuktha","middleName":"","lastName":"D","suffix":""},{"id":536685012,"identity":"886f8869-8ed2-49b9-b0ff-ce7da6014a61","order_by":1,"name":"Chandravel Saravanan","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Chandravel","middleName":"","lastName":"Saravanan","suffix":""},{"id":536685013,"identity":"bab6f6bf-683d-4c33-b971-6c389dac589b","order_by":2,"name":"Anusha Jayasimhan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0ElEQVRIiWNgGAWjYJACZgYGNjkIkw2MiNNiTLIWhsQGmBaCgH92j/Hnghq+9Pn9Zw8wfCg7zMAn3YBfi8SdM2bSM46x5W64kZfAOOPcYQY2mQMErLmRY8bMwwbUIsFjwMzbBtQikYBfh/yNHOPPPP/Y0uX7zxgw/yVGi8GNHANp3ja2BIYDOQbMjMRoMbyRVibN28dmuAGo92DPuXQeglrkbiRv/szz7Zg80GGGD36UWcvJzyCgBQqOgckDQMxDlHogqCFW4SgYBaNgFIxEAAAcHDuAazFIPwAAAABJRU5ErkJggg==","orcid":"","institution":"","correspondingAuthor":true,"prefix":"","firstName":"Anusha","middleName":"","lastName":"Jayasimhan","suffix":""}],"badges":[],"createdAt":"2025-10-29 06:28:53","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-7976248/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7976248/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":94729028,"identity":"bedf7597-ead9-44e4-8535-7a487c9dad66","added_by":"auto","created_at":"2025-10-30 07:04:30","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":479726,"visible":true,"origin":"","legend":"","description":"","filename":"EfficientPromptCompressiononEdgeDevices.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7976248/v1_covered_2d12ce02-7a0d-491b-b486-c10009465e9a.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eEfficient Prompt Compression on Edge Devices\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":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":"Large language models, Prompt compression, edge devices","lastPublishedDoi":"10.21203/rs.3.rs-7976248/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7976248/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLarge Language Models (LLMs) like GPT-4, BERT, and DeBERTa are widely used for tasks such as question answering, text summarization, and reasoning in many domains [1]–[3]. However, running these models on small or low-power devices such as mobile phones and IoT systems is challenging because they require large amounts of memory, processing power, and time when handling long text inputs called prompts. To solve this problem, researchers have developed several prompt compression methods that shorten prompts while preserving their key meaning. These include summarization-based [4], embedding-based [5], and graph-based reasoning techniques [6]. Recent methods such as LLMLingua-2 [15] and Prompt Compression with Context-Aware Sentence Encoding [20] have further improved compression quality while maintaining reasoning consistency and efficiency. Building upon these works, this paper proposes an Efficient Prompt Compression on Edge Devices framework that integrates embedding retrieval, causal-temporal reasoning, and coherence validation into a single, lightweight process. The framework produces interpretable reasoning graphs that retain only the most important information, enabling faster and more efficient processing. Experiments conducted on the CQR dataset demonstrate that the proposed model achieves high reasoning accuracy while significantly reducing computational cost, making it suitable for real-time deployment on edge devices. BLEU, and F1, making it suitable for real-world deployment on low-power devices.\u003c/p\u003e","manuscriptTitle":"Efficient Prompt Compression on Edge Devices","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-30 03:41:03","doi":"10.21203/rs.3.rs-7976248/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":"cdb01bc6-2a76-43cf-a158-653c5e82d9a0","owner":[],"postedDate":"October 30th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":57064942,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-10-30T03:41:04+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-30 03:41:03","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7976248","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7976248","identity":"rs-7976248","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}