Attention Amplification in Multilingual LLMs: Why Script Representation Matters

Attention Amplification in Multilingual LLMs: Why Script Representation Matters | 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 Attention Amplification in Multilingual LLMs: Why Script Representation Matters Yash Mishra, Suyash Mishra, Kedarnath senapati This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8959575/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 Modern Large Language Models (LLMs) inherently exhibit a profound architectural bias toward English and other Latin-script languages, inadvertently erecting a severe “script barrier” for the vast majority of the world’s linguistic diversity. This barrier stems primarily from the inefficient subword tokenization of non-Roman scripts, such as Devanagari, where standard algorithms aggressively fragment text into high-fertility sequences. This fragmentation not only drastically shrinks the effective context window but also quadratically amplifies the computational cost of self-attention. To circumvent this tokenization bottleneck, this paper investigates romanization—the transliteration of native scripts into the Latin alphabet—as a highly efficient computational bridge. By aligning the input representation with the pre-existing orthographic strengths of Englishcentric models, romanization serves as a pragmatic interface layer rather than a linguistic replacement, fundamentally mitigating the computational penalties imposed by standard tokenizers. Our comprehensive empirical analysis, integrating a primary case study with findings from the ROMANSETU framework, demonstrates that romanizing Hindi text yields a consistent 2.5x to 4x reduction in token count. This efficiency directly translates to competitive or superior performance across a wide array of Natural Language Understanding (NLU) and Natural Language Generation (NLG) tasks, particularly in generative and knowledge-retrieval domains. Furthermore, we formalize this computational overhead by deriving an attention amplification factor, revealing that native Devanagari processing requires over an order of magnitude more attention computation per unit of semantic content compared to its Romanized equivalent. We also systematically characterize the limitations of this pipeline, notably the risks of transliteration error propagation and the nuanced performance degradation on complex morpho-syntactic reasoning tasks. Ultimately, while romanization provides a powerful and immediately deployable strategy for enhancing multilingual AI efficiency, its necessity highlights the pressing, long-term requirement for fundamentally script-agnostic tokenization and multilingual model architectures. Large Language Models Romanization Subword Tokenization Devanagari Token Fertility Attention Amplification Multilingual NLP Indic NLP 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-8959575","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":596440224,"identity":"e6118dc9-4b6c-4e8a-bdbf-dbf795a5c64f","order_by":0,"name":"Yash Mishra","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA80lEQVRIiWNgGAWjYHACNoaPfyR4+OUPHwByJGSI0sI4s8FGTnIGWwJICw9RWph5G9KMDW7wGIB4hLUYHD/+7AHvjsOJDbd7Pr+6UWPBw8B++OgGvFrO5JgbSJ45nNg45+w265xjQIfxpKXdwKvlQA6bhAHb4cRmhtxtxkA2D9A7Zvi1nH/+TCIBqKWNIeeZcc4/YrTcSDCTONiWZswjkcP8OLeNCC2SN96YSTacsZGT4DlmxpzbJ8HDRsgvfOfTn0n/qZDgsT/e/Phzzrc6OX72w8fwalE4gGCzSYBJfMpBQL4BwWb+QEj1KBgFo2AUjEwAAKmlTQZsW6kyAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0002-5004-1797","institution":"Nitk Surathkal","correspondingAuthor":true,"prefix":"","firstName":"Yash","middleName":"","lastName":"Mishra","suffix":""},{"id":596440225,"identity":"2d306d2a-a02e-4018-8727-efb87f5b7025","order_by":1,"name":"Suyash Mishra","email":"","orcid":"","institution":"IIIT Manipur","correspondingAuthor":false,"prefix":"","firstName":"Suyash","middleName":"","lastName":"Mishra","suffix":""},{"id":596440226,"identity":"1eba7259-c1f8-463d-840f-d1cf023e41e7","order_by":2,"name":"Kedarnath senapati","email":"","orcid":"","institution":"Nitk Surathkal","correspondingAuthor":false,"prefix":"","firstName":"Kedarnath","middleName":"","lastName":"senapati","suffix":""}],"badges":[],"createdAt":"2026-02-24 16:33:40","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-8959575/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8959575/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103388573,"identity":"44a6f34e-3f76-4377-aa6e-ec00b1ca3fb6","added_by":"auto","created_at":"2026-02-25 07:13:38","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":503622,"visible":true,"origin":"","legend":"","description":"","filename":"snarticle.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8959575/v1_covered_f20eeb42-ef70-45d0-a9d7-2d98093918a2.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eAttention Amplification in Multilingual LLMs: Why Script Representation Matters\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":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":"Large Language Models, Romanization, Subword Tokenization, Devanagari, Token Fertility, Attention Amplification, Multilingual NLP, Indic NLP","lastPublishedDoi":"10.21203/rs.3.rs-8959575/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8959575/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eModern Large Language Models (LLMs) inherently exhibit a profound architectural bias toward English and other Latin-script languages, inadvertently erecting a severe “script barrier” for the vast majority of the world’s linguistic diversity. This barrier stems primarily from the inefficient subword tokenization of non-Roman scripts, such as Devanagari, where standard algorithms aggressively fragment text into high-fertility sequences. This fragmentation not only drastically shrinks the effective context window but also quadratically amplifies the computational cost of self-attention. To circumvent this tokenization bottleneck, this paper investigates romanization—the transliteration of native scripts into the Latin alphabet—as a highly efficient computational bridge. By aligning the input representation with the pre-existing orthographic strengths of Englishcentric models, romanization serves as a pragmatic interface layer rather than a linguistic replacement, fundamentally mitigating the computational penalties imposed by standard tokenizers.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOur comprehensive empirical analysis, integrating a primary case study with findings from the ROMANSETU framework, demonstrates that romanizing Hindi text yields a consistent 2.5x to 4x reduction in token count. This efficiency directly translates to competitive or superior performance across a wide array of Natural Language Understanding (NLU) and Natural Language Generation (NLG) tasks, particularly in generative and knowledge-retrieval domains. Furthermore, we formalize this computational overhead by deriving an attention amplification factor, revealing that native Devanagari processing requires over an order of magnitude more attention computation per unit of semantic content compared to its Romanized equivalent. We also systematically characterize the limitations of this pipeline, notably the risks of transliteration error propagation and the nuanced performance degradation on complex morpho-syntactic reasoning tasks. Ultimately, while romanization provides a powerful and immediately deployable strategy for enhancing multilingual AI efficiency, its necessity highlights the pressing, long-term requirement for fundamentally script-agnostic tokenization and multilingual model architectures.\u003c/p\u003e","manuscriptTitle":"Attention Amplification in Multilingual LLMs: Why Script Representation Matters","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-25 07:10:55","doi":"10.21203/rs.3.rs-8959575/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":"a00d7c71-a66d-4dc7-ad16-9b75e7488a2d","owner":[],"postedDate":"February 25th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-25T07:10:55+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-25 07:10:55","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8959575","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8959575","identity":"rs-8959575","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}