Reasoning Distillation by Prompt Optimization

Reasoning Distillation by Prompt Optimization | 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Share Download PDF Article Reasoning Distillation by Prompt Optimization Marcin Koralewski This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8231090/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 12 You are reading this latest preprint version Abstract Large Language Models (LLMs) increasingly rely on explicit chains of thoughts (CoT) reasoning to produce reliable and interpretable outputs. However, transferring such reasoning behaviour from one system to another typically requires expensive dataset construction, full-model retraining, or large task-specific corpora used for distillation. In this work, we introduce a lightweight prompt-level distillation framework that aligns a student model with the reasoning patterns of an external 'parent' source ? which may be a larger model or a human expert. Instead of constructing a curated supervision dataset, our method operates solely on reasoning traces generated by the parent. These traces serve as optimization targets for an automated prompt-search procedure that improves the logical consistency and step-wise reasoning of the student without modifying its parameters. Across multiple reasoning benchmarks, we show that prompt-level distillation substantially narrows the performance gap between student and parent models while eliminating the cost of dataset preparation and model training. This approach provides a practical pathway for disseminating high-quality reasoning behaviours in settings where computational resources, data availability, or human labor are limited. Physical sciences/Engineering Physical sciences/Mathematics and computing prompt optimization reasoning knowledge transfer Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 08 May, 2026 Reviews received at journal 20 Apr, 2026 Reviewers agreed at journal 20 Apr, 2026 Reviewers agreed at journal 01 Apr, 2026 Reviewers agreed at journal 03 Mar, 2026 Reviews received at journal 26 Feb, 2026 Reviewers agreed at journal 16 Feb, 2026 Reviewers invited by journal 02 Dec, 2025 Editor invited by journal 02 Dec, 2025 Editor assigned by journal 29 Nov, 2025 Submission checks completed at journal 29 Nov, 2025 First submitted to journal 28 Nov, 2025 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-8231090","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":553964842,"identity":"f3dcb236-9714-491d-a67f-a0098cca00ce","order_by":0,"name":"Marcin Koralewski","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA2klEQVRIiWNgGAWjYFACNhBhAWYc+ECCFgkw4+AMkrUw8xCjQbeBLfFxRY2EnMH5Y4mHbf7YyBncSGD88DEHtxazA2yHDc8ckzA2uJF24HAOTxqQkcAsOXMbPi3sbZINbBKJ226wNxzOkTicOHNGAhszL0Et/yTqt50/3nDYwoAoLWzHJBvbJBLMDgAdxpBwOLFfgpCWw2zJho19Eob7b6QlHOw5kGbMz/OwGb9fjrcZPmz4ZiMv2X/M+MMPYIixsScf/PARjxYGZkwhxgY86kfBKBgFo2AUEAMAU4ZQMJr4BaQAAAAASUVORK5CYII=","orcid":"","institution":"Zoovu","correspondingAuthor":true,"prefix":"","firstName":"Marcin","middleName":"","lastName":"Koralewski","suffix":""}],"badges":[],"createdAt":"2025-11-28 14:08:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8231090/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8231090/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97348655,"identity":"443e2858-3b6d-40b4-9794-b1ebf6e99ee7","added_by":"auto","created_at":"2025-12-03 12:01:52","extension":"json","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3201,"visible":true,"origin":"","legend":"","description":"","filename":"37d84914b5f04838a36baadc0e7b472e.json","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/b61667fc164dcd2da504ef19.json"},{"id":97348659,"identity":"2cc3601c-1ce0-4d78-a120-2c530b31b915","added_by":"auto","created_at":"2025-12-03 12:01:52","extension":"xml","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":68582,"visible":true,"origin":"","legend":"","description":"","filename":"37d84914b5f04838a36baadc0e7b472e1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/6a500954ccf58dfd561b592c.xml"},{"id":97348657,"identity":"795893ce-3c33-4f51-bf0b-0cd2f1f0ccb9","added_by":"auto","created_at":"2025-12-03 12:01:52","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":730708,"visible":true,"origin":"","legend":"","description":"","filename":"KoralewskiRevisedReasoning.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/86088d7cb511ca9aa04b30ad.pdf"},{"id":97371479,"identity":"9c9eb0d2-5f0f-43a7-bee2-170c5ff67cf3","added_by":"auto","created_at":"2025-12-03 16:29:02","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":163616,"visible":true,"origin":"","legend":"","description":"","filename":"accuracy41barplot.png","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/5d207a2ae240a35cf81af123.png"},{"id":97348662,"identity":"0a419315-37a0-4b55-ba91-c2d3eac3b357","added_by":"auto","created_at":"2025-12-03 12:01:53","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":241038,"visible":true,"origin":"","legend":"","description":"","filename":"accuracyo3barplot.png","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/e1986598aba53e5887677802.png"},{"id":97371126,"identity":"91716c37-ee8b-48a6-bb3c-76802b71d716","added_by":"auto","created_at":"2025-12-03 16:28:25","extension":"odt","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16422,"visible":true,"origin":"","legend":"","description":"","filename":"coverletterkoralewski.odt","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/1a6271155fad0a04555a3d58.odt"},{"id":97370782,"identity":"2620a33a-6040-4f86-9045-69e913f19543","added_by":"auto","created_at":"2025-12-03 16:27:55","extension":"cls","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":55857,"visible":true,"origin":"","legend":"","description":"","filename":"snjnl.cls","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/879b4d0fc321981c63ab3b9c.cls"},{"id":97348660,"identity":"598d6baf-2768-4063-a2e0-b40307eea8cd","added_by":"auto","created_at":"2025-12-03 12:01:52","extension":"bst","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":64164,"visible":true,"origin":"","legend":"","description":"","filename":"snmathphysnum.bst","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/a9e08b722b6c7f993664332a.bst"},{"id":97348663,"identity":"1541170c-00c2-452c-96eb-fe57f46d498e","added_by":"auto","created_at":"2025-12-03 12:01:53","extension":"xml","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":73695,"visible":true,"origin":"","legend":"","description":"","filename":"37d84914b5f04838a36baadc0e7b472e1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/6487aebe89b6ea9695f61c3c.xml"},{"id":97370890,"identity":"0e5f454d-03f1-462b-a866-e45e596f10f1","added_by":"auto","created_at":"2025-12-03 16:28:06","extension":"html","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":79073,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1/e0891a60c7c4a9d55a6dc6fe.html"},{"id":97373051,"identity":"f0822876-1334-4537-ae42-cd4adbf015e3","added_by":"auto","created_at":"2025-12-03 16:33:56","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":693327,"visible":true,"origin":"","legend":"","description":"","filename":"KoralewskiRevisedReasoning.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8231090/v1_covered_b566ecb4-6e9b-4582-95eb-3b8105465668.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Reasoning Distillation by Prompt Optimization","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":"prompt optimization, reasoning, knowledge transfer","lastPublishedDoi":"10.21203/rs.3.rs-8231090/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8231090/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLarge Language Models (LLMs) increasingly rely on explicit chains of thoughts (CoT) reasoning to produce reliable and interpretable outputs. However, transferring such reasoning behaviour from one system to another typically requires expensive dataset construction, full-model retraining, or large task-specific corpora used for distillation. In this work, we introduce a lightweight prompt-level distillation framework that aligns a student model with the reasoning patterns of an external 'parent' source ? which may be a larger model or a human expert. Instead of constructing a curated supervision dataset, our method operates solely on reasoning traces generated by the parent. These traces serve as optimization targets for an automated prompt-search procedure that improves the logical consistency and step-wise reasoning of the student without modifying its parameters. Across multiple reasoning benchmarks, we show that prompt-level distillation substantially narrows the performance gap between student and parent models while eliminating the cost of dataset preparation and model training. This approach provides a practical pathway for disseminating high-quality reasoning behaviours in settings where computational resources, data availability, or human labor are limited.\u003c/p\u003e","manuscriptTitle":"Reasoning Distillation by Prompt Optimization","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-03 12:01:48","doi":"10.21203/rs.3.rs-8231090/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"20832173629617397406087768802759392696","date":"2026-05-08T10:46:51+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-20T22:38:27+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"236724291972012644774033221243203606967","date":"2026-04-20T04:21:41+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"77909961610759847237603381804820806513","date":"2026-04-01T10:37:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"18215555073699108445839332451017224555","date":"2026-03-03T10:47:24+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-27T04:57:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"160671906560530785731274575997430169175","date":"2026-02-16T19:49:08+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-02T06:29:36+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-12-02T05:31:24+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-11-29T06:21:18+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-11-29T06:19:54+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-11-28T13:52:26+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":"e5aa82c8-98de-4c36-b2be-51363acbf743","owner":[],"postedDate":"December 3rd, 2025","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"20832173629617397406087768802759392696","date":"2026-05-08T10:46:51+00:00","index":248,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":58945503,"name":"Physical sciences/Engineering"},{"id":58945504,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2025-12-03T12:01:48+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-03 12:01:48","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8231090","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8231090","identity":"rs-8231090","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}