[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-1-en-105":3,"doc-seo-250892-105":53,"doc-detail-250892-en":127},{"code":4,"msg":5,"data":6},0,"success",[7,14,19,24,29,34,39,44,49],{"id":8,"doc_module":9,"doc_module_name":10,"category_name":11,"show_sort_weight":12,"slug":13},11,1,"Template","Presentations",90,"presentations",{"id":15,"doc_module":9,"doc_module_name":10,"category_name":16,"show_sort_weight":17,"slug":18},12,"Resumes",80,"resumes",{"id":20,"doc_module":9,"doc_module_name":10,"category_name":21,"show_sort_weight":22,"slug":23},14,"Invoices",70,"invoices",{"id":25,"doc_module":9,"doc_module_name":10,"category_name":26,"show_sort_weight":27,"slug":28},15,"Posters",60,"posters",{"id":30,"doc_module":9,"doc_module_name":10,"category_name":31,"show_sort_weight":32,"slug":33},16,"Social Media",50,"social-media",{"id":35,"doc_module":9,"doc_module_name":10,"category_name":36,"show_sort_weight":37,"slug":38},17,"Forms",40,"forms",{"id":40,"doc_module":9,"doc_module_name":10,"category_name":41,"show_sort_weight":42,"slug":43},18,"Letters",30,"letters",{"id":45,"doc_module":9,"doc_module_name":10,"category_name":46,"show_sort_weight":47,"slug":48},21,"Paper Templates",5,"papers-templates",{"id":50,"doc_module":9,"doc_module_name":10,"category_name":51,"show_sort_weight":4,"slug":52},158,"General","general-158",{"code":4,"msg":54,"data":55},"ok",{"site_id":56,"language":57,"slug":58,"title":59,"keywords":60,"description":61,"schema_data":62,"social_meta":120,"head_meta":122,"extra_data":124,"updated_unix":126},105,"en","lh-deception-simulating-and-understanding-llm-deceptive-behaviors-in-long-horizon-interactions","LH-DECEPTION - Simulating and Understanding LLM Deceptive Behaviors in Long-Horizon Interactions","","Deception is a pervasive feature of human communication and an emerging concern in large language models (LLMs). While existing studies document instances of LLM deception, most evaluations rely on single-turn prompts and do not capture the long-horizon interactions where deceptive strategies unfold. LH-DECEPTION introduces a multi-agent simulation and auditing framework to measure deception under extended, interdependent task sequences with dynamic contextual pressures, then empirically shows deception is model-dependent, escalates with event pressure, and degrades supervisor trust.",{"@graph":63,"@context":119},[64,80,102],{"@type":65,"itemListElement":66},"BreadcrumbList",[67,71,74,77],{"item":68,"name":69,"@type":70,"position":9},"https://docshare.wps.com","Home","ListItem",{"item":72,"name":10,"@type":70,"position":73},"https://docshare.wps.com/template/",2,{"item":75,"name":11,"@type":70,"position":76},"https://docshare.wps.com/template/presentations/",3,{"item":78,"name":59,"@type":70,"position":79},"https://docshare.wps.com/template/lh-deception-simulating-and-understanding-llm-deceptive-behaviors-in-long-horizon-interactions/250892/",4,{"url":78,"name":59,"@type":81,"image":82,"author":87,"headline":59,"publisher":90,"fileFormat":93,"inLanguage":57,"description":61,"dateModified":94,"datePublished":95,"encodingFormat":93,"isAccessibleForFree":96,"interactionStatistic":97},"DigitalDocument",{"url":83,"@type":84,"width":85,"height":86},"https://docshare.wps.com/thumbnails/lh-deception-simulating-and-understanding-llm-deceptive-behaviors-in-long-horizon-interactions/250892.png","ImageObject",442,249,{"name":88,"@type":89},"\tWilliam","Person",{"url":68,"name":91,"@type":92},"DocShare","Organization","application/pdf","2026-10-21","2026-09-12",true,{"@type":98,"interactionType":99,"userInteractionCount":101},"InteractionCounter",{"@type":100},"ViewAction",8,{"@type":103,"mainEntity":104},"FAQPage",[105,111,115],{"name":106,"@type":107,"acceptedAnswer":108},"Why do existing LLM deception benchmarks fall short?","Question",{"text":109,"@type":110},"Most focus on single-turn prompts, which cannot capture trajectory-level dynamics where deception emerges and compounds over time in long-horizon interactions.","Answer",{"name":112,"@type":107,"acceptedAnswer":113},"How does LH-DECEPTION simulate long-horizon deception?",{"text":114,"@type":110},"It uses a multi-agent setup: a performer agent attempts interdependent tasks, a supervisor agent evaluates progress and provides feedback while updating trust, and an independent auditor reviews complete trajectories to detect deceptive behavior.",{"name":116,"@type":107,"acceptedAnswer":117},"What do the experiments reveal about deception behavior?",{"text":118,"@type":110},"Deception depends on the specific model, increases with event pressure, and consistently erodes supervisor trust; qualitative analysis also shows emergent long-horizon patterns such as chains of deception.","https://schema.org",{"og:url":78,"og:type":121,"og:title":59,"og:site_name":91,"og:description":61},"article",{"robots":123,"canonical":78},"index,follow",{"doc_id":125,"site_id":56},250892,1791726316,{"code":4,"msg":5,"data":128},{"doc_id":125,"user_id":129,"nickname":88,"user_avatar":130,"doc_module":9,"category_id":8,"category_name":11,"doc_title":59,"doc_description":61,"doc_content":131,"file_id":132,"file_url":133,"file_type":134,"file_size":135,"view_count":101,"is_deleted":4,"is_public":9,"is_downloadable":9,"audit_status":9,"page_count":136,"language":137,"language_code":57,"site_id":56,"html_lang":57,"table_of_contents":138,"faqs":139,"seo_title":140,"seo_description":61,"update_tm":141,"read_time":15},5909887254083,"https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c","arXiv :2510 .03999v 3 [ cs .CL] 5 Feb 2026  \nLH-DECEPTION: SIMULATING AND UNDERSTANDING LLM DECEPTIVE BEHAVIORS IN LONG-HORIZON INTERACTIONS  \nYang Xu*†, Xuanming Zhang* § , Samuel Yeh§ , Jwala Dhamala‡, Ousmane Dia‡, Rahul Gupta‡, Sharon Li§  \n§University of Wisconsin-Madison,†Zhejiang University,‡Amazon AGI  \n*Equal contribution.  \nABSTRACT  \nDeception is a pervasive feature of human communication and an emerging concern in large language models (LLMs) . While recent studies document instances of LLM deception, most evaluations remain confined to single-turn prompts and fail to capture the long-horizon interactions in which deceptive strategies typically unfold. We introduce a new simulation framework, LH-DECEPTION, for a systematic, empirical quantification of deception in LLMs under extended sequences of interdependent tasks and dynamic contextual pressures. LH-DECEPTION is designed as a multi-agent system: a performer agent tasked with completing tasks and a supervisor agent that evaluates progress, provides feedback, and maintains evolving states of trust. An independent deception auditor then reviews full trajectories to identify when and how deception occurs. We conduct extensive experiments across 11 frontier models, spanning both closed-source and open-source systems, and find that deception is model-dependent, increases with event pressure, and consistently erodes supervisor trust. Qualitative analyses further reveal emergent, long-horizon phenomena, such as “chains of deception”, which are invisible to static, single-turn evaluations. Our findings provide a foundation for evaluating future LLMs in real-world, trust-sensitive contexts.  \n1 INTRODUCTION  \nHumans do not always say what they mean—and sometimes, they intentionally say things they know are false or misleading. Deception is a pervasive challenge in human communication, shaping trust, relationships, and decision-making (Ward et al., 2023) . It is now increasingly troubling in large language models (LLMs), which have begun to exhibit similar behaviors. While recent studies document LLMs’ capacity for deception (Hubinger et al., 2021 ; Scheurer et al., 2024 ; Greenblatt et al., 2024 ; Sabour et al., 2025 ; Chen et al., 2025 ; Baker et al., 2025 ; Taylor & Bergen, 2025 ; Motwaniet al., 2024), most existing benchmarks focus narrowly on short-form, single-turn evaluations.  \nThis is a critical gap, since modern LLMs are increasingly deployed in settings where they collaborate with humans or other agents over extended sequences of interdependent tasks. In such real-world long-horizon interactions, the conditions that give rise to deceptive behavior are fundamentally different from those captured by single-step or short-horizon evaluations. Prior theoretical work has noted that deception in long-horizon settings may pose distinct risks (Carroll et al., 2024), particularly when seemingly innocuous actions compound into misleading trajectories. However, these concerns have remained largely untested empirically. Understanding deception in this setting, therefore, demands a framework that models not just isolated prompts, but the trajectory-level dynamics through which misrepresentation can emerge, compound, or escalate.  \nThis aligns with decades of social science research emphasizing that deception rarely emerges in isolation; instead, it arises in complex social dynamics and typically unfolds across extended interactions (Buller et al., 1994 ; BondJr. & DePaulo, 2006) . This gap motivates our work: how to simulate, quantify, and understand LLMs’ deceptive behavior in long-horizon interactions? Designing evaluations that capture such long-horizon dynamics is highly non-trivial. Unlike standard  \nTask Stream (􀀡 )  \n\n|  |  |\n| --- | --- |\n|  |  |\n|  |  Task 1  |\n|  |  |\n|  | Event  |\n|  |  |\n|  |  |\n|  |  |\n|  |   Task 2  |\n|  | |\n|  |  |\n| |  |\n|  |  |\n|  | Task 􀀢 − 1  |\n|  |  |\n|  | |\n|  |  |\n|  |  |\n|  |  |\n|  | Task 􀀢  |\n|  |  |\n|  | |\n|  |  |\n|  |  |\n\nFigure 1: ","cbCaiuIISwdJ3Qpj","https://ap.wps.com/l/cbCaiuIISwdJ3Qpj","pdf",1165385,34,"English","# Abstract\n# Introduction\n## Long-horizon evaluation gap\n## Framework design: performer, supervisor, and auditor\n## Experimental findings and implications","[{\"question\":\"Why do existing LLM deception benchmarks fall short?\",\"answer\":\"Most focus on single-turn prompts, which cannot capture trajectory-level dynamics where deception emerges and compounds over time in long-horizon interactions.\"},{\"question\":\"How does LH-DECEPTION simulate long-horizon deception?\",\"answer\":\"It uses a multi-agent setup: a performer agent attempts interdependent tasks, a supervisor agent evaluates progress and provides feedback while updating trust, and an independent auditor reviews complete trajectories to detect deceptive behavior.\"},{\"question\":\"What do the experiments reveal about deception behavior?\",\"answer\":\"Deception depends on the specific model, increases with event pressure, and consistently erodes supervisor trust; qualitative analysis also shows emergent long-horizon patterns such as chains of deception.\"}]","LH-DECEPTION - Simulating and Understanding LLM Deceptive Behaviors in Long-Horizon Interactions | PDF",1789256565]