From Persona to Personalization: A Survey on Role-Playing Language Agents

• RPLAs simulate historical figures, fictional characters, and everyday individuals, enabling digital clones, chatbot characters, role-playing games, and social science research.Source: §1 · p. 2
• Recent LLM advances enable RPLAs to reproduce personas’ knowledge, language, behavior, and personalities through prompted contexts or parametric knowledge.RPLAs also provide testing grounds for studying LLM social capabilities and sources of synthetic training data.Source: §1 · p. 2
• The survey categorizes personas as Demographic, Character, and Individualized, progressing from statistical group stereotypes to recognized figures and continuously updated personal profiles.The taxonomy’s scope includes definitions, data sources, examples, use cases, and applications.Source: §1 · pp. 2–3, Figure 1 · p. 2
• These persona types can coexist within one RPLA, while individualized personas emphasize dynamic learning from user interactions and real-world activities.The survey uses a Socrates philosophy tutor as an example combining demographic, character, and individualized personas.Source: §1 · p. 3
• The survey systematically reviews RPLA methodologies, persona-specific research, risks including toxicity and bias, limitations, future directions, and current products.Its organization covers background, general RPLA research, demographic, character, and individualized personas, risks, conclusions, and an applications-oriented product survey.Source: §1 · p. 3

2.1 The Roadmap of Large Language Models

• LLMs’ in-context learning and instruction following enable RPLAs to absorb persona information, follow role-playing prompts, and imitate characters without parameter updates (Brown et al., 2020; Ouyang et al., 2022).Step-by-step reasoning and social intelligence further enrich human likeness and emotional support in RPLA applications (Wei et al., 2022b; Wang et al., 2024b; Sclar et al., 2023; Light et al., 2023).Source: §2.1 · pp. 3–4
• LLMs exhibit human-like traits including self-awareness, values, emotional perception, psychopathy, and personalities, which Shanahan et al. (2023) attribute to role-playing rather than consciousness.These traits contribute to the compelling sense of human likeness associated with LLM-based RPLAs (Park et al., 2023).Source: §2.1 · pp. 3–4
• Retrieval-augmented generation reduces factual errors by retrieving knowledge during inference, while extended context lets LLMs process novels and documents without fragmenting persona information (Karpukhin et al., 2020; Lewis et al., 2020).Both approaches expand the reliability and scope of role-playing scenarios (Shao et al., 2023; Chen et al., 2023c; Zhou et al., 2023a).Source: §2.1 · p. 4

2.2 LLM-powered Language Agents

• Planning modules decompose complex tasks and adapt actions using strategies such as CoT and ReAct, enabling realistic, dynamic RPLA interactions in games and social simulations (Wei et al., 2022b; Yao et al., 2023b).These capabilities address long-horizon tasks by incorporating feedback from the environment (Rana et al., 2023; Yuan et al., 2023).Source: §2.2 · p. 4
• Tool use gives RPLAs access to APIs, knowledge bases, external models, and customized actions, improving accuracy and contextual appropriateness in specialized or complex scenarios.Tools enable interaction with environments such as games and software applications beyond the LLM’s intrinsic capabilities (Patil et al., 2023; Zhuang et al., 2024; Wang et al., 2023a).Source: §2.2 · pp. 4–5
• Memory mechanisms store agent profiles, relationships, experiences, user data, and environmental context, enabling consistent behavior, continuity, personalization, and more relevant responses over time.Short-term memory resides within the transformer context window, whereas long-term memory is commonly stored in vector or natural-language databases (Shinn et al., 2023).Source: §2.2 · p. 5

3.1 RPLA Definition

• RPLAs are categorized as demographic, character, and individualized personas, progressing from broad demographic groups to profiles customized through individual behavioral and preference data.Demographic personas support group-specific simulations and tasks; character personas represent real or fictional figures for entertainment and emotional engagement; individualized personas adapt dynamically for personalized assistants, companions, or proxies.Source: §3.1 · pp. 5, 7

3.2 RPLA Construction

• RPLAs simulate intricate personas from descriptive narratives, dialogues, historical behaviors, and extensive textual materials, using either parametric training or nonparametric prompting.Parametric training includes pre-training, supervised fine-tuning, and reinforcement learning, while prompting supplies persona descriptions, demonstrations, instructions, and increasingly memory modules for retrieving traits or past interactions.Source: §3.2 · pp. 7–8
• Parametric learning generally targets demographic information and well-known characters, whereas prompting is commonly used for fictional and highly personalized characters.Current development also includes demographic steering, foundation models, agent frameworks, and persona profiles for specified individuals.Source: §3.2 · p. 8

3.3 RPLA Evaluation

• RPLA evaluation distinguishes role-playing capability from persona fidelity, assessing general abilities such as conversation, engagement, consistency, emotion understanding, theory of mind, and problem-solving alongside knowledge, language, personality, beliefs, and decisions.Evaluation methods include automatic evaluation with or without ground truth, multiple-choice questions, and human-based evaluation; current systems remain below fully human-level agents and nuanced evaluation remains limited.Source: §3.3 · p. 8

4.1 Definition

• Demographic personas represent groups defined by shared occupations, interests, or personality types through distinctive language, expertise, behavior, viewpoints, and social complexity.They translate broad RPLA capabilities into virtual representations of demographic archetypes and support social experiments or complex tasks involving multiple demographic agents.Source: §4.1 · p. 9

4.2 Analysis of Demographics

• RPLAs possess inherent personality, political, economic, and ethical demographics, can be assessed with human psychological tests, and can shift behavior when directly prompted into specified roles.These characteristics are statistically consistent yet differ across LLMs; role assignment can diverge from inherent traits and enable diverse human-like simulations, but may produce toxic or biased outputs through stereotype amplification or jailbreaking.Source: §4.2 · pp. 9–10

4.3 Application of Demographics

• Demographic personas improve specialized single-agent responses and support cooperative problem-solving in multi-agent systems, including complex reasoning, mathematical modeling, and software development.These roles can increase response length, depth, quality, contextual appropriateness, and engagement, while frameworks such as ChatDev and MetaGPT distribute specialized responsibilities across agents.Source: §4.3 · p. 10
• Diversified demographic agents simulate social behavior across games and daily interactions, with role variation, population scale, and interaction producing realistic or unplanned collective behaviors.RPLAs have matched or surpassed human performance in diplomacy-focused games, contributed to social simulations, and revealed emergent dynamics in larger multi-agent environments.Source: §4.3 · pp. 10–11

5.1 Definition

• Character RPLAs portray established figures, including celebrities, historical figures, and fictional characters, while research increasingly targets faithful reproduction of linguistic style, knowledge, personality, and decision-making.Character understanding covers character prediction and personality understanding, including identifying characters, modeling relationships, forecasting actions, and decoding traits from dialogue and behavior.Source: §5.1 · p. 11

5.2 Data for Character RPLAs

• Character data combines descriptions of identities, relationships, and traits with demonstrations of linguistic, cognitive, and behavioral patterns; descriptions are foundational, while demonstrations improve vividness and fidelity.Demonstrations may come from experience extraction, dialogue synthesis, or human annotation, but available datasets cover few characters; synthesized dialogue depends on teacher-LLM quality and often requires filtering.Source: §5.2 · pp. 11–13

5.3 Construction of Character RPLAs

• Character RPLAs are constructed through parametric training or nonparametric prompting, with long-term memory increasingly retrieving relevant character and interaction data when context limits prevent including everything.Parametric methods use pre-training or supervised fine-tuning, whereas prompting directly supplies character data in context; ongoing user interactions can personalize characters and create point-in-time role-playing challenges.Source: §5.3 · pp. 13–14, §5.2 · p. 13

5.4 Evaluation of Character RPLAs

• Evaluation separates character-independent capabilities from character fidelity, measuring engagement, conversation quality, anthropomorphic abilities, style, knowledge, personality, and thinking through automatic, multiple-choice, and human methods.Ground-truth-free LLM evaluation is limited for unfamiliar characters and nuanced knowledge, while human evaluation is costly and difficult to reproduce; recent work increasingly evaluates original characters because established-character datasets are challenging.Source: §5.4 · p. 14

6.1 Definition

• Individualized personas tailor RPLAs to users’ needs, experiences, and preferences, supporting conversation, recommendation, and increasingly autonomous task solving.They typically embody digital clones or personal assistants and require persona data collection followed by persona modeling.Source: §6.1 · p. 16
• Personalized RPLAs still face challenges with long inputs, large search spaces, sparse and noisy interactions, domain-specific and multimodal understanding, privacy, ethics, and real-time response speed.These limitations affect both persona construction and deployment in seamless personalized applications.Source: §6.1 · p. 16

6.2 Data Collection of Individualized Persona

• Personalized-persona data comprises profiles, interactions, and domain knowledge across languages and modalities.Profiles encode explicit identity and characteristics, interactions capture evolving behavior, and domain knowledge improves understanding of user needs and persona consistency.Source: §6.2 · pp. 17–18

6.3 Modeling Individualized Persona

• Offline and online learning form the main modeling paradigms: offline methods commonly fine-tune on datasets, whereas online methods update personas from incoming user data.Offline approaches trade accuracy against efficiency and struggle with changing personas, while online approaches rely on memory management, retrieval, and sometimes interactive fine-tuning to handle sparse, noisy feedback.Source: §6.3 · p. 18

6.4 Evaluation for LLMs and Individualized Persona

• Evaluation follows the three application tiers: conversation tests persona understanding and style replication, recommendation tests preference-based ranking, and task solving tests personalized execution.Metrics and capabilities include response accuracy, fluency, similarity, recall, MRR, Top-k accuracy, theory of mind, tool usage, task automation, privacy, and external-tool interaction.Source: §6.4 · pp. 18–19

7 Risks Beneath RPLA Applications

• The survey organizes RPLA risks into toxicity, bias, hallucination, privacy violations, and technical challenges in real-world deployment.It also discusses anthropomorphism-related harms, including social isolation and manipulation of public opinion.Source: §7 · p. 19, §7.6 · p. 23

7.1 Toxicity

• RPLAs can amplify toxic content because role alignment may reproduce characters’ immoral behaviors, while fully safe role-playing can reduce training-data coverage and generalization.Prompt engineering and semantic censorship aim to reduce toxic outputs while preserving versatility and effectiveness.Source: §7.1 · p. 20

7.2 Bias

• RPLA bias includes reasoning, political, and role-based forms, while simplistic persona construction fails to capture nuanced, scenario-dependent behavior.Bias arises from training-data imbalances and user prompting; data cleaning, fairness-aware classifiers, and richer persona attributes are proposed mitigations.Source: §7.2 · pp. 20–21

7.3 Hallucination

• RPLA hallucination includes responses that conflict with assigned roles or time periods, such as characters knowing events outside their historical or narrative scope.Mitigation requires teaching agents to express ignorance or forget out-of-scope knowledge rather than relying only on external knowledge bases.Source: §7.3 · p. 21

7.4 Privacy Violations

• RPLAs may disclose private information or enable identity inference through attacks and individual-persona prompting, creating risks of identity theft and unauthorized access.Anonymization, privacy protocols, and leak-detection tools such as ProPILE are proposed safeguards.Source: §7.4 · pp. 21–22

7.5 Technical Challenges in Real-world Deployment

• Real-world RPLA deployment is constrained by weak social intelligence, ego-centric interaction, limited long-context reasoning, computational demands, immersion requirements, and knowledge gaps.These limitations can impair emotional support, persona consistency, complex role portrayal, and evaluation reliability.Source: §7.5 · p. 22

7.6 Anthropomorphism

• Over-reliance on anthropomorphized RPLAs may reduce human contact and social skills, worsen loneliness and related distress, or enable undisclosed manipulation of public opinion.These risks are especially concerning when systems spread misinformation during elections, public-health crises, or other sensitive contexts.Source: §7.6 · p. 23

Survey Contributions

• The survey classifies RPLAs into Demographic Persona, Character Persona, and Individualized Persona, while identifying risks and ethical concerns requiring focused mitigation.The taxonomy captures development from generically assigned personas to highly personalized ones and leaves substantial research and application opportunities.Source: §8 · p. 23

Future Directions: Causal Data Analysis for Decision-making

• RPLAs should analyze causal factors underlying experiences rather than merely mimic observable actions, enabling more justifiable and informed decisions.Extracting and confirming causality from intertwined experiences requires advanced analytics and deeper data interpretation.Source: §8 · p. 23

Future Directions: Improved Decision-making

• RPLA decision-making should optimize outcomes for individual scenarios by using role context to avoid mistakes, resolve dilemmas, and potentially exhibit superhuman intelligence.This requires comprehensive collection and use of contextual intricacies associated with the roles.Source: §8 · p. 23

Future Directions: RPLA as Personal Assistants for Personal Decision-making

• RPLAs could become comprehensive personal assistants managing customized shopping, travel planning, and recommendations through multimodal data and advanced visualization.Such capabilities aim to improve personalization and efficiency in everyday tasks.Source: §8 · pp. 23–24

Future Directions: Social Simulation through Autonomous Role-Playing

• Autonomous role-playing can support social simulation experiments that examine psychological and sociological behavior, personality effects on social intelligence, and human interaction dynamics.RPLAs can act as versatile test subjects across diverse scenarios by role-playing different characters.Source: §8 · p. 24

A RPLA Products

• Existing RPLA products are categorized into persona-oriented and task-oriented applications, as listed in Table 4.These applications reflect how persona and personalization demands shape and propel RPLA research.Source: A RPLA Products · p. 43

Product scope

• Persona-oriented RPLAs primarily simulate specific fictional characters, historical figures, celebrities, or other defined personas across chatbots, game NPCs, and tailored use cases.Products are categorized by whether their primary interaction is between humans and RPLAs or among multiple RPLAs.Source: §A.1 · pp. 43–44

Human–RPLA interactions

• Human-facing products include established-character agents such as Meta AI, Hello History, and ChatFAI, alongside general or user-defined personas offered by Character.ai and Replika.Commercial systems typically use brief prompt-based character settings, whereas ChatHaruhi (Li et al., 2023a) and RoleLLM (Wang et al., 2024a) curate detailed character data but are less commonly adopted commercially for generality and cost efficiency.Source: §A.1 · p. 44
• Many human-facing RPLAs dynamically learn user prompts and preferences through interaction, often using long-term memory modules to evolve their personas.Some products reproduce a user’s “digital self,” including language, voice, or visual appearance, and support video presentations or conferences for sales, marketing, and customer service.Source: §A.1 · p. 44

RPLA–RPLA interactions

• Multi-RPLA products target interactive gaming and simulations in which users orchestrate storylines or role-play as characters interacting with generated agents.Systems such as Ememe.ai and AI Dungeon support sandbox-like settings, while SageRPG and Hidden Door use interacting characters or franchise-based worlds to expand actions beyond predefined storylines.Source: §A.1 · p. 44

• Task-oriented RPLAs use human-like communication and domain expertise to deliver personalized services across education, healthcare, human resources, customer service, content creation, real estate, shopping, fitness, travel, and finance.They are closely related to personalized RPLA research and are deployed in specialized AI products.Source: §A.2 · p. 44
• In education and human resources, RPLAs tailor learning content, recommendations, teaching materials, assessments, interview preparation, and career advice to individual users.Examples include Jagoda.AI, Khanmigo, Duolingo Max, Eduaide.Ai, Huru AI, and Careers AI.Source: §A.2 · p. 44
• Real-estate and content-generation products use user needs, preferences, interactions, and market or audience data to generate tailored recommendations, descriptions, creative media, and social-media strategies.These systems support property discovery, market insights, writing, image and video generation, audience analysis, summarization, and interactive question-answering.Source: §A.2 · pp. 44–45
• Healthcare and fitness agents personalize guidance by using patient or user data, adapting responses and plans through feedback, progress, abilities, biometric information, and physical characteristics.Healthcare agents support medical guidance and logistics, while fitness agents act as coaches that set goals, adapt exercises, and provide multimodal feedback.Source: §A.2 · pp. 45–46
• Travel, customer-service, and shopping agents personalize concierge services, complaint handling, product matching, and recommendations by interpreting queries, preferences, emotions, behavior, and conversational feedback.They can also analyze interaction data to anticipate needs, improve first-contact resolution, discover trends, and support business decisions.Source: §A.2 · p. 45
• Office-oriented RPLAs act as context-aware copilots by integrating workers’ documents, emails, code repositories, and other private files to generate content, answer requests, and streamline workflows.Microsoft 365 Copilot uses Microsoft Graph and continuous learning, while GitHub Copilot uses individual code repositories to support programmers.Source: §A.2 · p. 46
• Shopping Muse simulates in-store conversations to recommend products, match items, and identify trends based on user preferences.Source: §A.2 · p. 50
• Fitness-oriented agents provide advice, responses, and customized workout plans tailored to users’ fitness goals and plans.Source: §A.2 · p. 50
• Workplace RPLAs support productivity through Microsoft 365 integration, personalized coding assistance, and autonomous task execution.Source: §A.2 · p. 50