Two Tales of Persona in LLMs: A Survey of Role-Playing and Personalization

📝 Paper Summary

LLM Personalization LLM Role-Playing

This survey unifies disjoint research on LLM persona into two distinct streams: Role-Playing (LLMs acting as specific characters) and Personalization (LLMs adapting to user profiles), providing a comprehensive taxonomy and evaluation review.

Core Problem

Research on leveraging personas in LLMs is disorganized and lacks a systematic taxonomy, causing confusion between methods where LLMs adopt a persona versus methods where LLMs adapt to a user's persona.

Why it matters:

Current literature conflates 'persona' as a model attribute (role-playing) with 'persona' as a user attribute (personalization), hindering clear methodological development.
Lack of unified terminology makes it difficult for researchers to identify relevant prior work across fields like dialogue systems, recommendation, and agentic simulation.

Concrete Example: In a gaming context, an LLM might need to role-play a merchant (Role-Playing) while simultaneously adjusting its negotiation strategy based on the player's past purchasing history (Personalization). Without a clear taxonomy, these distinct adaptation mechanisms are often conflated.

Key Novelty

Unified Persona Taxonomy

Distinguishes 'LLM Role-Playing' (persona belongs to the LLM to adapt to an environment) from 'LLM Personalization' (persona belongs to the user to meet individualized needs).
Categorizes role-playing into emergent behaviors (voluntary, conformity, destructive) and schemas (single-agent vs. multi-agent).
Categorizes personalization across domains: Dialogue, Healthcare, Education, Search, and Recommendation.

Architecture

A conceptual diagram contrasting the two tales of persona: Role-Playing vs. Personalization.

Evaluation Highlights

Identifies that LLM personality evaluation relies heavily on psychological frameworks like Big Five and MBTI (Myers-Briggs Type Indicator).
Highlights PsychoBench as a key evaluation platform for assessing personality traits in LLMs.
Notes that role-playing evaluation often focuses on task success (e.g., code generation rates in MetaGPT), whereas personalization evaluation focuses on user satisfaction and alignment.

Breakthrough Assessment

7/10

While a survey paper (no new model proposed), it provides a critical structural definition to a chaotic field, clearly distinguishing between agentic role-play and user-centric personalization.

⚙️ Technical Details

Problem Definition

Setting: A survey and taxonomy definition for Persona-based LLM research.

Inputs: Existing literature on LLM Agents, Personalization, and Role-Playing

Outputs: Structured Taxonomy categorizing methods into LLM Role-Playing vs. LLM Personalization

Pipeline Flow

Literature Collection
Taxonomy Creation (Split: Role-Playing vs. Personalization)
Sub-categorization (Environment/Schema vs. Domain/Application)
Evaluation Method Analysis

System Modules

Taxonomy Branch 1: LLM Role-Playing (Categorization)

Classify methods where the LLM adopts a specific character/role

Model or implementation: N/A (Conceptual Framework)

Taxonomy Branch 2: LLM Personalization (Categorization)

Classify methods where the LLM adapts to a user's profile

Model or implementation: N/A (Conceptual Framework)

Evaluation Review

Review methods for assessing personality alignment

Model or implementation: N/A (Conceptual Framework)

Novel Architectural Elements

Unified View of Persona: Treating 'Role-Playing' and 'Personalization' as two sides of the same coin (Model-side vs. User-side adaptation).

Comparison to Prior Work

vs. General LLM Surveys: Specifically targets the concept of 'Persona' rather than general capabilities.
vs. Agent Surveys: Distinguishes user-centric personalization from agent-centric role-playing, whereas agent surveys typically focus only on the latter (task completion).
Novel contribution: First survey to explicitly dichotomize the field into 'Role-Playing' (Model Persona) and 'Personalization' (User Persona).

Limitations

The survey taxonomy relies on existing literature, which is rapidly evolving; new emergent behaviors may appear that fit neither category perfectly.
The distinction between role-playing and personalization can blur in mixed-initiative scenarios (e.g., a personalized role-playing tutor).
Does not propose a new metric or model, only organizes existing work.

Reproducibility

Code: https://github.com/MiuLab/PersonaLLM-Survey

The authors maintain a GitHub repository (https://github.com/MiuLab/PersonaLLM-Survey) containing the paper collection used for this survey. As this is a survey paper, no model training code is applicable, but the literature list is public.

📊 Experiments & Results

Evaluation Setup

Qualitative review and categorization of existing literature.

Benchmarks:

PsychoBench (Personality Evaluation)
PersonaChat (Dialogue Personalization)
MultiWoz (Task-oriented Dialogue)
MovieLens (Recommendation)

Metrics:

N/A (Survey paper)
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

Role-playing enhances LLM performance in specific domains (e.g., medical diagnosis via DR-CoT, software dev via ChatDev) by eliciting domain-specific parametric knowledge.
Multi-agent role-playing induces emergent social behaviors like conformity (aligning with group goals) and voluntary cooperation.
Personalization is domain-dependent: Dialogue focuses on consistency with user history; Search/RecSys focuses on ranking accuracy based on user logs; Education/Health focuses on empathy and user state tracking.
Evaluation of persona is bifurcated: Role-playing is evaluated by task success (did the 'doctor' diagnose correctly?), while personalization is evaluated by user alignment (did the 'tutor' explain it in a way *I* understand?).

📚 Prerequisite Knowledge

Prerequisites

Basic understanding of Large Language Models (LLMs)
Familiarity with Agentic AI concepts (agents, environments)
Basic knowledge of personalization (user profiles, history)

Key Terms

LLM Role-Playing: LLMs are tasked to play assigned personas (roles) and act based on environmental feedback to adapt to the environment (e.g., acting as a doctor or a game NPC).

LLM Personalization: LLMs are tasked to take care of user personas (background, history) to meet individualized needs (e.g., personalized search or tutoring).

Single-Agent Schema: One agent acts independently to achieve goals without necessary assistance from others (e.g., Voyager in Minecraft).

Multi-Agent Schema: Agents must collaborate or communicate to achieve goals (e.g., ChatDev for software development).

Voluntary Behavior: Agents proactively assisting peers or contributing resources without being explicitly forced, enhancing team efficiency.

Conformity Behavior: Agents adjusting their deviating decisions to align with team goals after receiving criticism.

Big Five: A taxonomy of personality traits: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism.

MBTI: Myers-Briggs Type Indicator, a self-report inventory designed to identify a person's personality type, strengths, and preferences.

CoT: Chain-of-Thought—a prompting technique where the model generates intermediate reasoning steps.