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

TL;DR

This survey classifies persona-based adaptations in LLMs into two streams: role-playing (LLMs adopt fixed personas within defined environments) and personalization (LLMs tailor outputs to user personas). It provides a unified taxonomy covering environments, schemas, and emergent behaviors, along with evaluation approaches for LLM personality. The paper highlights challenges such as building general frameworks, handling long-context personas, data and benchmark gaps, biases, and safety/privacy concerns, offering directions and a living collection to guide future work. Collectively, the work aims to standardize understanding and accelerate progress in persona-aware LLM research and applications.

Abstract

The concept of persona, originally adopted in dialogue literature, has re-surged as a promising framework for tailoring large language models (LLMs) to specific context (e.g., personalized search, LLM-as-a-judge). However, the growing research on leveraging persona in LLMs is relatively disorganized and lacks a systematic taxonomy. To close the gap, we present a comprehensive survey to categorize the current state of the field. We identify two lines of research, namely (1) LLM Role-Playing, where personas are assigned to LLMs, and (2) LLM Personalization, where LLMs take care of user personas. Additionally, we introduce existing methods for LLM personality evaluation. To the best of our knowledge, we present the first survey for role-playing and personalization in LLMs under the unified view of persona. We continuously maintain a paper collection to foster future endeavors: https://github.com/MiuLab/PersonaLLM-Survey

Figures (4)

• Figure 1: In Role-Playing, LLMs act according to assigned personas (i.e., roles) under a defined environment. For example, given role names with descriptions, LLMs role-play in a social simulation game. For Personalization, LLMs consider user personas to generate tailored responses to the same question. Dashed rectangles are prompts and solid rectangles are LLMs' responses.
• Figure 2: The taxonomy of LLM role-playing and LLM personalization (representative works shown only).
• Figure 3: An illustration of four LLM role-playing environments: Software Development\ref{['subsub:2-software']}, Game\ref{['subsub:2-games']}, Medical Application\ref{['subsub:2-medical']}, and LLM as Evaluators\ref{['subsub:2-evaluator']}. For each environment, we provide a simple scenario with a task description (red-bordered) and relevant personas (i.e., roles; blue-bordered). The dashed rectangle represents an example LLM role-playing prompt template. In addition to the above environments, past research also proposes general frameworks applicable to different environments \ref{['subsub:2-general']}.
• Figure 4: An illustration of five types of personalized LLMs: Recommendation\ref{['subsub:3-rec']}, Search\ref{['subsub:3-search']}, Education\ref{['subsub:3-education']}, Healthcare\ref{['subsub:3-healthcare']}, and Dialogue\ref{['subsub:3-diag']}. On the left side, dashed rectangles are prompts, and solid rectangles are the responses of LLMs. On the right side, we depict multi-turn interactions between LLMs and users.