Source-linked AI summary
From Persona to Personalization: A Survey on Role-Playing Language Agents
Jiangjie Chen, Xintao Wang, Rui Xu, Siyu Yuan, Yikai Zhang, Wei Shi, Jian Xie, Shuang Li, Ruihan Yang, Tinghui Zhu, Aili Chen, Nianqi Li, Lida Chen, Caiyu Hu, Siye Wu, Scott Ren, Ziquan Fu, Yanghua Xiao
TL;DR
RPLAs raise the question of how LLMs can simulate diverse personas while supporting useful applications and managing risks. This survey organizes the field’s methods and applications through a three-level persona taxonomy, then synthesizes its capabilities, risks, and limitations. It concludes that RPLAs span demographic, character, and individualized uses, but still face hallucination and social-intelligence challenges.
Problem
RPLAs require systematic understanding of how LLMs simulate diverse personas, support applications, and present risks and limitations.
Method
The survey reviews RPLA literature, establishes taxonomies, categorizes personas into three types, and examines data, construction, evaluation, applications, risks, and future directions.
Results
The survey identifies demographic, character, and individualized personas spanning specialized tasks, social simulations, entertainment, emotional engagement, personalized assistance, and digital clones.
Takeaways & Limitations
RPLA research and applications are organized by increasing personalization, from statistical group archetypes and established characters to continuously updated individual profiles.
Takeaways & Limitations
RPLAs remain challenged by character hallucination and insufficient social intelligence or theory of mind for socially intelligent interaction.
Abstract
from arXiv · showhide
Recent advancements in large language models (LLMs) have significantly boosted the rise of Role-Playing Language Agents (RPLAs), i.e., specialized AI systems designed to simulate assigned personas. By harnessing multiple advanced abilities of LLMs, including in-context learning, instruction following, and social intelligence, RPLAs achieve a remarkable sense of human likeness and vivid role-playing performance. RPLAs can mimic a wide range of personas, ranging from historical figures and fictional characters to real-life individuals. Consequently, they have catalyzed numerous AI applications, such as emotional companions, interactive video games, personalized assistants and copilots, and digital clones. In this paper, we conduct a comprehensive survey of this field, illustrating the evolution and recent progress in RPLAs integrating with cutting-edge LLM technologies. We categorize personas into three types: 1) Demographic Persona, which leverages statistical stereotypes; 2) Character Persona, focused on well-established figures; and 3) Individualized Persona, customized through ongoing user interactions for personalized services. We begin by presenting a comprehensive overview of current methodologies for RPLAs, followed by the details for each persona type, covering corresponding data sourcing, agent construction, and evaluation. Afterward, we discuss the fundamental risks, existing limitations, and future prospects of RPLAs. Additionally, we provide a brief review of RPLAs in AI applications, which reflects practical user demands that shape and drive RPLA research. Through this work, we aim to establish a clear taxonomy of RPLA research and applications, and facilitate future research in this critical and ever-evolving field, and pave the way for a future where humans and RPLAs coexist in harmony.
1 Introduction
The survey frames RPLAs as LLM-based systems that simulate assigned personas and organizes them into demographic, character, and individualized types. It reviews their methodologies, risks, limitations, future prospects, and applications.
- RPLA Motivation: Recent LLM advances support RPLAs that replicate historical figures, fictional characters, and everyday individuals in interactive applications.Examples include digital clones, AI chatbot characters, and role-playing video games.
- Persona Taxonomy: Demographic personas represent groups sharing characteristics and leverage statistical stereotypes embedded in LLMs.Examples include occupations, ethnic groups, and personality types.
- Persona Taxonomy: Character personas represent recognized celebrities, historical figures, or fictional characters using knowledge from model parameters or supplied contexts.Their role-playing depends on understanding curated materials associated with established characters.
- Persona Taxonomy: Individualized personas are continuously updated digital profiles built from personal data to support services such as digital clones and personal assistance.They emphasize individual experiences, needs, preferences, dynamic learning, and interaction with real-world activities.
- Persona Taxonomy: The three persona types can coexist progressively, as illustrated by an RPLA portraying Socrates as a personal philosophy tutor.The example combines demographic, character, and user-developed individualized aspects.
- Survey Scope: The survey systematically reviews RPLA literature, establishes methodological taxonomies, and examines persona categories, risks, limitations, future directions, and products.Its product review aims to connect theoretical insights with practical applications.
2 Preliminary
RPLAs build on emerging LLM abilities that support human-like role-playing, while agent modules add planning, tool use, memory, and retrieved information. These capabilities remain subject to hallucination and social-intelligence limitations.
- LLM Foundations: In-context learning, instruction following, step-by-step reasoning, and social intelligence provide foundations for complex LLM role-playing behavior.These abilities help models learn persona information, follow role instructions, reason through tasks, and support human-like interaction.
- LLM Foundations: LLMs exhibit human-like traits including self-awareness, values, emotional perception, psychopathy, and personalities, motivating anthropomorphic study.The passage presents these traits as emerging areas of research rather than established human equivalence.
- Agent Components: Retrieval-augmented generation dynamically retrieves knowledge during inference to enhance LLM capability and mitigate factually incorrect content.RAG integrates external data retrieval into the generative process.
- Agent Components: Planning modules decompose complex long-horizon tasks into subtasks and select actions using strategies such as CoT and ReAct with environmental feedback.The module adapts subsequent actions based on feedback from the environment.
- Agent Components: Tool-usage modules let agents execute actions through APIs, knowledge bases, external models, and application-specific tools.For RPLAs, these tools support interaction with environments such as games and software applications.
- Agent Components: Memory mechanisms store agent profiles and environmental information for future actions.Profiles can include basic information, psychological traits, and social relationships.
3 Overview of RPLAs
The survey defines RPLAs, distinguishes three persona categories by increasing specificity, and organizes construction methods around parametric training and nonparametric prompting. It also separates capability evaluation from persona fidelity evaluation.
- Persona Taxonomy: The taxonomy progresses from demographic personas to character personas and individualized personas as personalization increases.Demographic personas represent aggregated groups, character personas represent established figures, and individualized personas represent specific users.
- Demographic Persona: Demographic RPLAs use fictional archetypes derived from LLM pretraining and can be created efficiently with simple prompts for group simulations or specialized tasks.An example prompt is “You are a mathematician.”
- Character Persona: Character RPLAs use biographies, novels, and films to represent public figures or fictional entities for entertainment and emotional engagement.Typical applications include AI chatbots and virtual video-game characters.
- Individualized Persona: Individualized RPLAs adapt to continuously evolving behavioral and preference data to provide customized services as assistants, companions, or proxies.Their data may include personal profiles, dialogues, actions, and behaviors.
- RPLA Definition: RPLAs simulate intricate personas using profiles, dialogues, historical behaviors, and extensive textual materials such as books.Persona data supplies the narratives and behavioral information used to construct role-playing agents.
- Construction Methods: RPLA construction generally uses parametric training or nonparametric prompting, with parametric learning tending toward demographics and known characters and prompting toward fictional or highly personalized personas.The two approaches may also contribute concurrently to development.
- Evaluation: Evaluation distinguishes role-playing capability from persona fidelity, covering dimensions such as anthropomorphic ability, attractiveness, usefulness, conversation, and engagement.Capability evaluation targets foundation models and construction frameworks independently of specific personas.
4 Demographic Persona
Demographic RPLAs model group-level characteristics through language, knowledge, behavior, and viewpoints, while their assigned personas can improve task performance and simulate social behavior. However, demographic role-playing may also amplify stereotypes, biases, and toxicity.
- Demographic personas represent groups through characteristic language preferences, domain vocabulary, viewpoints, and behavioral nuances.
- 4.2 Analysis of Demographics: RPLAs can exhibit inherent personality, political, economic, and ethical characteristics that vary across language models.
- 4.2 Analysis of Demographics: Researchers can assess inherent demographic characteristics using human psychological instruments such as the Big Five Personality Test.
- 4.2 Analysis of Demographics: Persona assignment can also increase toxic or biased outputs when personas amplify stereotypes and training-data biases, with jailbreaks potentially bypassing safety mechanisms.
- 4.3 Application of Demographics: Assigning specific demographics often improves downstream performance, especially on tasks requiring persona-linked specialized knowledge and complex zero-shot reasoning.
- 4.3 Application of Demographics: In multi-agent settings, demographic personas support cooperative problem-solving and the simulation of collective social behaviors across games and broader environments.
5 Character Persona
Character RPLAs simulate established or original characters by combining character data with LLM capabilities for faithful role-playing. The survey organizes their data sources, construction methods, and evaluation around character understanding and fidelity.
- Definition: Character RPLAs represent recognized public figures, historical figures, fictional characters, and occasionally original characters.
- Character Understanding: Effective character role-playing requires models to understand characters, including their identities, relationships, traits, personalities, and likely behaviors.
- Data for Character RPLAs: Character data primarily comprises descriptions and demonstrations: descriptions provide foundational attributes, while demonstrations illustrate patterns that should generalize beyond exact examples.
- Limitations and Future Directions: Character-RPLA research remains constrained by limited datasets covering only a small selection of characters, while interaction data and point-in-time role-playing remain underexplored.
- Data for Character RPLAs: Character datasets are built through experience extraction, dialogue synthesis, or human annotation, trading off fidelity, scalability, background requirements, and labor cost.
- Construction: Construction uses parametric training or nonparametric prompting, with long-term memory modules retrieving character knowledge and interaction data when context becomes extensive.
- Evaluation: Evaluation separates character-independent foundation-model capabilities from persona fidelity for specific characters, covering role-playing abilities, conversational skills, and aligned behaviors.
6 Individualized Persona(lization)
Individualized persona RPLAs model users’ needs, experiences, preferences, and behavior for increasingly complex personalized services. Their central pipeline collects persona data and models it offline or online, while facing challenges from noisy, changing, multimodal, and privacy-sensitive user information.
- Definition: Personalized RPLAs emulate individual users by adapting to their unique needs, experiences, preferences, and behaviors.
- Applications: Applications span three tiers: personalized conversation, recommendation, and autonomous agents for complicated task solving.
- Applications: Task-solving personalized RPLAs support coding, travel planning, and research surveying while integrating with personal data, devices, services, and external software.
- Data and Modeling: Building individualized personas involves persona data collection followed by persona modeling, with data varying across formats, content, modalities, applications, and tasks.
- Challenges: Personalized RPLAs face challenges in processing long inputs, sparse and noisy interactions, domain-specific and multimodal contexts, privacy, ethics, and changing user personas.
- Persona Data: Individualized personas commonly use profiles, interactions, and domain knowledge; profiles describe users, interactions capture evolving behavior, and domain knowledge supports profile and interaction understanding.
- Persona Modeling: Offline learning trains on pre-existing datasets, whereas online learning updates personas with incoming data through memory, context management, or interactive fine-tuning.
7 Risks Beneath RPLA Applications
RPLAs face risks spanning toxic and biased outputs, hallucinated role-inconsistent information, privacy violations, and deployment challenges involving social intelligence and theory of mind.
- Toxicity: RPLAs can intensify toxicity because role-playing may align outputs with characters whose behavior violates societal moral standards.Safe general role-playing remains difficult because human-generated data contain toxic content, complicating construction of clean training corpora.
- Mitigation Strategies: Mitigation strategies include prompt engineering, semantic censorship, data cleaning, fairness-aware classifiers, anonymization, privacy protocols, and leak-detection tools.These approaches aim to reduce harmful outputs and privacy risks while preserving versatility and effectiveness.
- Bias: Bias in RPLAs includes reasoning, political, and role-based forms, creating tension between character authenticity and ethical standards.Role-based bias may arise when assigned personas express prejudicial views that conflict with ethical expectations.
- Hallucination: Character hallucination occurs when agents provide information inconsistent with assigned roles, including knowledge unavailable at a character’s historical point in time.For example, Shakespeare should not know about World War II, while a young Harry Potter should not mention future events.
- Privacy Violations: RPLA privacy violations may expose personal information or identify individuals from limited attributes, creating risks of identity theft and unauthorized access.The survey highlights both extraction attacks and the ability to infer identities from details such as location, gender, and birth date.
- Technical Challenges in Real-world Deployment: Current LLM limitations in social intelligence and theory of mind hinder adequate emotional support and can produce ego-centric conversational behavior.These abilities involve perceiving and reasoning about users’ emotions, beliefs, intentions, and needs.
8 Closing Remarks
The survey organizes RPLA research around a progression from demographic and character personas toward individualized personalization. It concludes that future systems should support more reasoned decisions, personal assistance, and social simulation while addressing identified risks.
- Closing Remarks: The survey classifies RPLA personas into Demographic, Character, and Individualized Persona types, representing progressively more personalized systems.The categories can coexist, and the classification describes a developmental trajectory from generic assignments to highly personalized personas.
- Future Directions on RPLA Systems: Future RPLAs should reason and make decisions that resemble or transcend assigned roles rather than merely mimic them.The survey identifies this capability as central to moving from persona-assigned role-playing toward personalization.
- Future Directions on RPLA Systems: Causal data analysis is proposed to support justifiable decisions by identifying underlying causes rather than only reproducing observable actions.Extracting and confirming causal factors from intertwined experiences remains challenging.
- Future Directions on RPLA Systems: Improved decision-making should tailor choices to individual scenarios, including avoiding mistakes and handling difficult dilemmas.The proposed direction allows decisions that may display advanced or superhuman intelligence.
- Future Directions on RPLA Systems: RPLAs may develop into personal assistants that manage Internet activities such as shopping, travel planning, and recommendation tasks.The survey also points to multimodal data handling and visualization as components of this direction.
- Future Directions on RPLA Systems: Autonomous role-playing could support social simulations that examine how personality traits influence social intelligence and human interaction dynamics.RPLAs would serve as role-playing subjects across diverse psychological and sociological scenarios.
A RPLA Products
RPLA products are grouped into persona-oriented and task-oriented categories, reflecting how user demands for persona and personalization shape RPLA research.
- RPLA Products: Existing RPLA products are distinguished as persona-oriented or task-oriented according to their primary application focus.The survey presents these two categories as a review of recent RPLA application trends.
A.1 Persona-oriented RPLA Products
Persona-oriented RPLA products support interactions with established or user-defined personas, dynamic adaptation through user interactions, and representations of users’ digital selves.
- Persona-oriented RPLA Products: Persona-oriented RPLAs commonly role-play fictional characters, historical figures, celebrities, or people with specified professions or personalities.They are used in entertainment applications such as chatbots and game NPCs, and may be forked for individual preferences.
- Interactions between Humans and RPLAs: Human–RPLA interactions can begin from established characters or develop through ongoing user interactions.Products such as Character.ai support both forms of interaction.
- Interactions between Humans and RPLAs: Users can create RPLAs with user-defined personas, typically using prompts that briefly describe character settings.Some research projects instead curate detailed character data for specific well-known characters.
- Interactions between Humans and RPLAs: Many products dynamically evolve personas by learning from user prompts and preferences, often through long-term memory modules.This enables the persona to adjust during interaction rather than remain fixed at initialization.
- Interactions between Humans and RPLAs: Digital-self RPLAs reproduce users’ language and sometimes physical characteristics such as voice or visual appearance.These systems support text chats, video presentations, and conferences for sales, marketing, and customer service.
- Interactions among RPLAs: Multi-agent RPLA products target interactive gaming and simulations in which users orchestrate storylines or play pivotal characters.Users may design simulation settings and characters with or without participating directly as players.
A.2 Task-oriented RPLAs
Task-oriented RPLAs act as personalized domain specialists across education, healthcare, human resources, customer service, content creation, real estate, shopping, fitness, travel, and finance. They use user profiles, behavior histories, and files to tailor assistance, recommendations, and workflows.
- Overview: Task-oriented RPLAs communicate human-like and provide personalized services as domain experts across many specialized application domains.The survey covers education, healthcare, human resources, customer service, content creation, real estate, shopping, fitness, travel, and finance.
- Education: Education agents adapt learning content and recommendations to learners while helping educators create teaching materials and assessments.Examples include Jagoda.AI, Khanmigo, Duolingo Max, Squirrel AI, and Eduaide.ai.
- Human resources: Human-resource agents tailor career support and interview preparation to job seekers’ profiles and interests.Applications provide interview answers, career advice, customized preparation materials, and practice using actual job listings.
- Real estate and content creation: Real-estate and content-generation agents produce personalized recommendations or creative materials, including property listings, market analysis, text, images, audio, and video.Real-estate systems analyze user needs and market data, while content tools tailor outputs to styles, themes, scenes, and objectives.
- Healthcare and fitness: Healthcare and fitness agents personalize guidance using patient or user data, while fitness coaches adapt exercises, goals, training plans, and feedback.Fitness platforms can use biometric information, physical characteristics, and fitness levels to tailor training plans.
- Service and productivity applications: Customer-service, travel, shopping, and office agents use preferences, interactions, context, or private files to provide tailored responses, recommendations, and productivity assistance.Examples include digital concierge services, product matching, customer support, and copilot assistance over documents and code repositories.