Dual-Path Perceptual Networks: A Stochastic Parallel Branch Architecture for Robust Reasoning in Large Language Models

Dual-Path Perceptual Networks: A Stochastic Parallel Branch Architecture for Robust Reasoning in Large Language Models | 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 Article Dual-Path Perceptual Networks: A Stochastic Parallel Branch Architecture for Robust Reasoning in Large Language Models fu chang lin This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8876183/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Large language models (LLMs) demonstrate remarkable linguistic fluency, yet remain fundamentally brittle in tasks demanding compositional reasoning, robustness to adversarial distractions, and grounded factual generation. A core architectural limitation underpinning this shortfall is their monolithic deterministic processing pipeline: conventional LLMs rely on a single feed-forward stream that greedily commits to a single representational interpretation at each layer, discarding alternative hypotheses critical for robust inference. Here we present Dual-Path Perceptual Networks (DPPN) , a lightweight plug-and-play architectural extension that equips Transformer-based LLMs with a parallel stochastic representational branch. This branch is explicitly engineered to encode two foundational inductive biases absent in standard LLMs: 1. Sustained high output entropy, achieved via multi-headed dropout that preserves representational uncertainty through network depth; 2. A non-adaptive feature basis, realized through fixed overcomplete random projections coupled with a trainable down-projection, yielding implicit regularization against overfitting. Both components are implemented exclusively with classical neural network modules and require no specialized hardware. The original deterministic Transformer stream and the stochastic branch are adaptively fused per token via a lightweight gating network that learns context-dependent weighting. We systematically evaluate DPPN across three challenging benchmarks: arithmetic reasoning with adversarial distractors, long-context multi-hop question answering, and factual precision in open-ended generation. Experiments on GPT-2 (124M–774M) and LLaMA-7B (via parameter-efficient fine-tuning) show that DPPN consistently outperforms the base LLM and parameter-matched baselines, including wider feed-forward networks, model ensembles, and parallel trainable MLPs. On the newly introduced Math-Distract benchmark, DPPN reduces reasoning errors by 41% relative to GPT-2 Large. On TruthfulQA, it improves factual precision by 7.9 percentage points. Ablation studies verify that fixed random projections, multi-headed stochasticity, and adaptive gating are individually indispensable to the observed performance gains. Our findings establish that structured stochasticity and non-adaptive representational constraints serve as powerful inductive biases for enhancing the robustness and reasoning capacity of LLMs, using only minimal architectural modifications. Physical sciences/Engineering Physical sciences/Mathematics and computing large language models robust reasoning stochastic regularization random projection parallel neural architectures Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 20 Apr, 2026 Reviewers invited by journal 18 Apr, 2026 Editor assigned by journal 18 Apr, 2026 Editor invited by journal 28 Mar, 2026 Submission checks completed at journal 25 Mar, 2026 First submitted to journal 25 Mar, 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-8876183","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":629521876,"identity":"7633fc7c-dd8b-453d-88dd-18f358d3b514","order_by":0,"name":"fu chang lin","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAwklEQVRIiWNgGAWjYBAC+/b2ww8+VPyTk2dvIFKLAc+ZNMMZZw4YG/YcIFaLRIKBNGfbgUSGGwlEajHnOZBgzHDmTgLjzMcbbzDU2EQT1GLZ3njgcUHFszx26bRiC4ZjabkNBPWcAdoy4wxzMePsHDMJxobDRGi5AfQLbxtzYsPNM0RqMYBoOZzYcIOHSC2SPeBATgMGMtAvCcT4hZ8dHJU2wKg8vPHGhxobIvyC7EiJBFKUQ7SQqmMUjIJRMApGBgAAfm5HvIeTaxEAAAAASUVORK5CYII=","orcid":"","institution":"","correspondingAuthor":true,"prefix":"","firstName":"fu","middleName":"chang","lastName":"lin","suffix":""}],"badges":[],"createdAt":"2026-02-14 01:53:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8876183/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8876183/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107913468,"identity":"e6df72cd-a9ba-4407-9d26-2145ba9195af","added_by":"auto","created_at":"2026-04-27 13:45:45","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":214529,"visible":true,"origin":"","legend":"","description":"","filename":"DualPathPerceptualNetworksUnleashingRobustReasoninginLLMs.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8876183/v1_covered_26866810-24b6-4f73-aaf6-80a33b608d60.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Dual-Path Perceptual Networks: A Stochastic Parallel Branch Architecture for Robust Reasoning in Large Language Models","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":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"large language models, robust reasoning, stochastic regularization, random projection, parallel neural architectures","lastPublishedDoi":"10.21203/rs.3.rs-8876183/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8876183/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLarge language models (LLMs) demonstrate remarkable linguistic fluency, yet remain fundamentally brittle in tasks demanding compositional reasoning, robustness to adversarial distractions, and grounded factual generation. A core architectural limitation underpinning this shortfall is their monolithic deterministic processing pipeline: conventional LLMs rely on a single feed-forward stream that greedily commits to a single representational interpretation at each layer, discarding alternative hypotheses critical for robust inference.\u003c/p\u003e \u003cp\u003eHere we present \u003cb\u003eDual-Path Perceptual Networks (DPPN)\u003c/b\u003e, a lightweight plug-and-play architectural extension that equips Transformer-based LLMs with a parallel stochastic representational branch. This branch is explicitly engineered to encode two foundational inductive biases absent in standard LLMs:\u003c/p\u003e \u003cp\u003e1. Sustained high output entropy, achieved via multi-headed dropout that preserves representational uncertainty through network depth;\u003c/p\u003e \u003cp\u003e2. A non-adaptive feature basis, realized through fixed overcomplete random projections coupled with a trainable down-projection, yielding implicit regularization against overfitting.\u003c/p\u003e \u003cp\u003eBoth components are implemented exclusively with classical neural network modules and require no specialized hardware. The original deterministic Transformer stream and the stochastic branch are adaptively fused per token via a lightweight gating network that learns context-dependent weighting.\u003c/p\u003e \u003cp\u003eWe systematically evaluate DPPN across three challenging benchmarks: arithmetic reasoning with adversarial distractors, long-context multi-hop question answering, and factual precision in open-ended generation. Experiments on GPT-2 (124M\u0026ndash;774M) and LLaMA-7B (via parameter-efficient fine-tuning) show that DPPN consistently outperforms the base LLM and parameter-matched baselines, including wider feed-forward networks, model ensembles, and parallel trainable MLPs. On the newly introduced Math-Distract benchmark, DPPN reduces reasoning errors by 41% relative to GPT-2 Large. On TruthfulQA, it improves factual precision by 7.9 percentage points. Ablation studies verify that fixed random projections, multi-headed stochasticity, and adaptive gating are individually indispensable to the observed performance gains.\u003c/p\u003e \u003cp\u003eOur findings establish that structured stochasticity and non-adaptive representational constraints serve as powerful inductive biases for enhancing the robustness and reasoning capacity of LLMs, using only minimal architectural modifications.\u003c/p\u003e","manuscriptTitle":"Dual-Path Perceptual Networks: A Stochastic Parallel Branch Architecture for Robust Reasoning in Large Language Models","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-27 13:45:37","doi":"10.21203/rs.3.rs-8876183/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"5483593751833898957225888178286811799","date":"2026-04-20T17:05:14+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-18T10:06:53+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-18T10:04:18+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-03-28T21:08:48+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-25T06:05:07+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-03-25T06:00:26+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"dc91fb8d-5b71-4772-b13b-26a52acf4246","owner":[],"postedDate":"April 27th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":67008261,"name":"Physical sciences/Engineering"},{"id":67008262,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-04-27T13:45:37+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-27 13:45:37","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8876183","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8876183","identity":"rs-8876183","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}