Training compact language models for artificial emotional intelligence: from bluffing to trust in a social deduction game

Training compact language models for artificial emotional intelligence: from bluffing to trust in a social deduction game | 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 Training compact language models for artificial emotional intelligence: from bluffing to trust in a social deduction game Christian Poglitsch, Johanna Pirker This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7348060/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 Social awareness is essential for effective interpersonal communication and informed decision-making, particularly in interactive and high-stakes environments. Emotional intelligence, the ability to recognize, understand, and regulate one’s own emotions while accurately interpreting and responding to the emotions of others serves as a foundational component of social competence. This is especially critical in social deduction games, where players must navigate strategic deception, manage trust, and infer intentions through nuanced psychological cues. This study explores how large language models (LLMs) can mimic emotionally intelligent behavior in the context of Blood on the Clocktower, a dialogue-based social deduction game with complex and dynamic game states. In such games, players must skillfully decide when to reveal information, bluff, or mislead others, making mastery of both the game mechanics and social dynamics essential. We task an LLM with inferring the hidden game state solely through conversation with other players and selecting ac-tions—including dialogue, game-related, and role-specific decisions—based on its evolving understanding. We use GPT-4o to generate high-quality training data and serve as a benchmark for evaluating performance. A smaller model, Mistral-7B-Instruct-v0.3, is first trained on this data and subsequently self-trained using Monte Carlo Tree Search guided interactions. We show that small LLMs can achieve competitive performance in social deduction settings by leveraging minimal but well-structured training data. Binary-branch MCTS proves sufficient for enabling models to find winning strategies. The trained Mistral-7B-Instruct-v0.3 model was able to outperform GPT-4o in our evaluation. This result suggests that reinforcement-guided retraining can provide a scalable and effective pathway for developing models that mimic emotional intelligence in social deduction games, particularly in environments where nuanced dialogue and subtle social cues render manual prompt engineering ineffective or impractical. Physical sciences/Mathematics and computing Biological sciences/Psychology Social science/Psychology Full Text Additional Declarations No competing interests reported. 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-7348060","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":509501503,"identity":"3fddbf97-8006-4901-9d8c-95e6676a99fb","order_by":0,"name":"Christian Poglitsch","email":"data:image/png;base64,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","orcid":"","institution":"Graz University of Technology, Institute of Human-Centred Computing","correspondingAuthor":true,"prefix":"","firstName":"Christian","middleName":"","lastName":"Poglitsch","suffix":""},{"id":509501504,"identity":"77b1c84c-686c-42ba-b6e0-d19e99ee9f96","order_by":1,"name":"Johanna Pirker","email":"","orcid":"","institution":"Technical University of Munich","correspondingAuthor":false,"prefix":"","firstName":"Johanna","middleName":"","lastName":"Pirker","suffix":""}],"badges":[],"createdAt":"2025-08-11 15:53:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7348060/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7348060/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102297778,"identity":"47bc0be2-2a10-4bae-beed-522bd69bacc0","added_by":"auto","created_at":"2026-02-10 10:29:09","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":266722,"visible":true,"origin":"","legend":"","description":"","filename":"BloodontheClocktowerEmotionalIntelligence7.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7348060/v1_covered_71910bdd-fc69-4dce-be01-6c10d965cab8.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Training compact language models for artificial emotional intelligence: from bluffing to trust in a social deduction game","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":"","lastPublishedDoi":"10.21203/rs.3.rs-7348060/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7348060/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSocial awareness is essential for effective interpersonal communication and informed decision-making, particularly in interactive and high-stakes environments. Emotional intelligence, the ability to recognize, understand, and regulate one’s own emotions while accurately interpreting and responding to the emotions of others serves as a foundational component of social competence. This is especially critical in social deduction games, where players must navigate strategic deception, manage trust, and infer intentions through nuanced psychological cues. This study explores how large language models (LLMs) can mimic emotionally intelligent behavior in the context of Blood on the Clocktower, a dialogue-based social deduction game with complex and dynamic game states. In such games, players must skillfully decide when to reveal information, bluff, or mislead others, making mastery of both the game mechanics and social dynamics essential.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe task an LLM with inferring the hidden game state solely through conversation with other players and selecting ac-tions—including dialogue, game-related, and role-specific decisions—based on its evolving understanding. We use GPT-4o to generate high-quality training data and serve as a benchmark for evaluating performance. A smaller model, Mistral-7B-Instruct-v0.3, is first trained on this data and subsequently self-trained using Monte Carlo Tree Search guided interactions. We show that small LLMs can achieve competitive performance in social deduction settings by leveraging minimal but well-structured training data. Binary-branch MCTS proves sufficient for enabling models to find winning strategies. The trained Mistral-7B-Instruct-v0.3 model was able to outperform GPT-4o in our evaluation. This result suggests that reinforcement-guided retraining can provide a scalable and effective pathway for developing models that mimic emotional intelligence in social deduction games, particularly in environments where nuanced dialogue and subtle social cues render manual prompt engineering ineffective or impractical.\u003c/p\u003e","manuscriptTitle":"Training compact language models for artificial emotional intelligence: from bluffing to trust in a social deduction game","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-19 02:42:54","doi":"10.21203/rs.3.rs-7348060/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":"4c2de9fb-9add-4008-883e-7bb4919e8f2a","owner":[],"postedDate":"September 19th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":54112966,"name":"Physical sciences/Mathematics and computing"},{"id":54112967,"name":"Biological sciences/Psychology"},{"id":54112968,"name":"Social science/Psychology"}],"tags":[],"updatedAt":"2026-02-10T06:11:53+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-19 02:42:54","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7348060","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7348060","identity":"rs-7348060","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}