[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-293699-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-293699-en":131},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","multi-personality-generation-of-llms-at-decoding-time","Multi-Personality Generation of LLMs at Decoding-time","","Multi-personality generation for LLMs enables simultaneous embodiment of multiple personalization attributes, addressing users with interacting preferences. Existing retraining-based methods are expensive and hard to scale, while decoding-time approaches often depend on external models or heuristics, reducing flexibility and robustness. This work introduces a decoding-time Multi-Personality Generation (MPG) framework that reformulates the problem as sampling from a target strategy using implicit density ratios. Speculative chunk-level rejection sampling reduces overhead while preserving quality, with improvements up to 16%–18% on MBTI and role-playing benchmarks.",{"@graph":14,"@context":73},[15,34,56],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/multi-personality-generation-of-llms-at-decoding-time/293699/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/multi-personality-generation-of-llms-at-decoding-time/293699.png","ImageObject",300,407,{"name":42,"@type":43},"Eden","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-11-05","2026-09-17",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",12,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"What problem does Multi-Personality Generation (MPG) address?","Question",{"text":63,"@type":64},"It enables LLMs to embody multiple personalization attributes at the same time, supporting complex user preferences in applications like assistants and personalized QA.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"Why do existing retraining-based methods have drawbacks?",{"text":68,"@type":64},"They require high training costs and poor scalability when adapting to new or changing preferences, often needing full retraining.",{"name":70,"@type":61,"acceptedAnswer":71},"How does the proposed MPG framework work at decoding time?",{"text":72,"@type":64},"MPG avoids extra training by leveraging implicit density ratios in singledimensional models, reformulating generation as sampling from a target strategy. It implements this efficiently with speculative chunk-level rejection sampling to validate chunks in a sliding window.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},293699,1789667715,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,106,111,115,120,123,127],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":107,"doc_module":4,"doc_module_name":25,"category_name":108,"show_sort_weight":109,"slug":110},7,"Healthcare",40,"healthcare",{"id":112,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":113,"slug":114},8,30,"research-report",{"id":116,"doc_module":4,"doc_module_name":25,"category_name":117,"show_sort_weight":118,"slug":119},9,"Religion & Spirituality",20,"religion-spirituality",{"id":118,"doc_module":4,"doc_module_name":25,"category_name":121,"show_sort_weight":118,"slug":122},"World Cup","world-cup",{"id":124,"doc_module":4,"doc_module_name":25,"category_name":125,"show_sort_weight":124,"slug":126},10,"Lifestyle","lifestyle",{"id":128,"doc_module":4,"doc_module_name":25,"category_name":129,"show_sort_weight":97,"slug":130},19,"General","general",{"code":4,"msg":82,"data":132},{"doc_id":79,"user_id":133,"nickname":42,"user_avatar":134,"doc_module":4,"category_id":112,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":140,"language":141,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":12,"update_tm":80,"read_time":145,"preview_page_count":4},1374391974468,"https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0","Multi-Personality Generation of LLMs at Decoding-time  \nRongxin Chen† Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences Beijing, China [chenrongxin24s@ict.ac.cn](chenrongxin24s@ict.ac.cn)  \nYunFan Li† Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences Beijing, China [liyunfan24s@ict.ac.cn](liyunfan24s@ict.ac.cn)  \nYige Yuan† Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences Beijing, China [yuanyige20z@ict.ac.cn](yuanyige20z@ict.ac.cn)  \narXiv :2511 .01891v4 [ cs .CL] 15 Jan 2026  \nBingbing Xu*† Institute of Computing Technology, Chinese Academy of Sciences Beijing, China [xubingbing@ict.ac.cn](xubingbing@ict.ac.cn)  \nHuawei Shen† Institute of Computing Technology, Chinese Academy of Sciences Beijing, China [shenhuawei@ict.ac.cn](shenhuawei@ict.ac.cn)  \nAbstract  \nMulti-personality generation for LLMs, enabling simultaneous embodiment of multiple personalization attributes, is a fundamental challenge. Existing retraining-based approaches are costly and poorly scalable, while decoding-time methods often rely on external models or heuristics, limiting flexibility and robustness. In this paper, we propose a novel Multi-Personality Generation (MPG) framework under the decoding-time combination paradigm. It flexibly controls multi-personality without relying on scarce multi-dimensional models or extra training, leveraging implicit density ratios in singledimensional models as a \"free lunch\" to reformulate the task as sampling from a target strategy aggregating these ratios. To implement MPG efficiently, we design Speculative Chunk-level based Rejection sampling (SCR), which generates responses in chunks and parallelly validates them via estimated thresholds within a sliding window. This significantly reduces computational overhead while maintaining high-quality generation. Experiments on MBTI personality and Role-Playing demonstrate the effectiveness of MPG, showing improvements up to 16%–18% . Code and data are available at [https://github.com/Libra117/MPG](https://github.com/Libra117/MPG).  \n1 Introduction  \nIn recent years, Large Language Models (LLMs) [34] have witnessed rapid development, with parameter scales surging from billions to trillions and training data covering diverse domains such as text and code [4, 21], thereby endowing them with capabilities in reasoning and semantic understanding [12] . Within this context, multi-personality generation has emerged as a critical task, which requires LLMs to transform from general-purpose tools into personalized intelligences capable of understanding and addressing the nuanced needs of individual users, and is widely applied in scenarios such as personalized question answering, artificial assistants, and intelligent customer service [6, 10, 16, 21] .  \nMulti-personality generation requires generated text to simultaneously embody multiple, potentially interacting personalization  \n* Corresponding author.  \n†State Key Laboratory of AI Safety of Chinese Academy of Sciences (CAS) .  \nFigure 1: Role-Playing in Multi-personality generation.  \nattributes. Figure 1 illustrates multi-personality generation in roleplay, where the personality set specifies multi-dimensional requirements, and LLMs need to take all these dimensions into account when generating responses. However, the inherent complexity of modeling multiple personalizations concurrently, combined with the challenge of balancing and integrating these traits within generalpurpose LLMs, amplifies the difficulty of this research.  \nExisting studies can be broadly categorized into retraining-based and decoding-time methods. Retraining methods seek to encode multiple preference dimensions into a single model during training via multi-objective optimization. This involves techniques like multiobjective reinforcement learning where models learn to balance competing preferen","cbCaivFLxEYLYUzy","https://ap.wps.com/l/cbCaivFLxEYLYUzy","pdf",937658,14,"English","# Introduction\n## Multi-personality generation overview\n## Retraining-based methods\n## Decoding-time methods and their limitations\n## Decoding-time combination paradigm","[{\"question\":\"What problem does Multi-Personality Generation (MPG) address?\",\"answer\":\"It enables LLMs to embody multiple personalization attributes at the same time, supporting complex user preferences in applications like assistants and personalized QA.\"},{\"question\":\"Why do existing retraining-based methods have drawbacks?\",\"answer\":\"They require high training costs and poor scalability when adapting to new or changing preferences, often needing full retraining.\"},{\"question\":\"How does the proposed MPG framework work at decoding time?\",\"answer\":\"MPG avoids extra training by leveraging implicit density ratios in singledimensional models, reformulating generation as sampling from a target strategy. It implements this efficiently with speculative chunk-level rejection sampling to validate chunks in a sliding window.\"}]","Multi-Personality Generation of LLMs at Decoding-time | PDF",35]