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

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.

Introduction. The striking capabilities of large language models (LLMs), exemplified by ChatGPT (OpenAI, 2022), have significantly advanced the field of natural language processing (NLP; Wei et al., 2023; Madaan et al., 2024; Shinn et al., 2024). Recently, in addition to using LLMs as NLP task solvers or general-purpose chatbots, the question of how to adapt LLMs for specific context has received great attention. To this end, leveraging personas has resurfaced as an ideal lens for adapting LLMs in target scenarios (Chen et al., 2023a, 2024). By incorporating personas, LLMs can generate more contextually appropriate responses, maximizing their utility and effectiveness for specific applications. However, the growing literature on persona in the LLM era is relatively disorganized, lacking a unifying overview. In this paper, we aim to close the gap by offering a comprehensive survey and a systematic categorization of existing studies. Specifically, we divide current research into two main streams, namely LLM Role-Playing and LLM Personalization, as illustrated in Figure 1.

Discussion / Conclusion. As LLM personalization continues to advance in education domains, individuals could easily access personalized educational contents, lecture materials, and receive affordable tutoring, largely benefiting minority groups with limited resources. However, the concern of polarizing trends might arise, where the privileged group enjoys private tutors and underrepresented individuals only have access to LLM-powered supports (Li et al., 2023c). Also, personalized LLMs for healthcare could potentially be widely integrated into clinical scenarios, mental health assessments, or prescribed therapeutic treatments in the near future, where critical questions such as legal considerations of the liability associated with these personalized systems needs careful considerations (Swift and Allen, 2010). As discussed in (§4), though methods for LLM personality evaluation have been proposed, there still lacks a unifying understanding of how to quantify personality in LLMs (Fang et al., 2023). Song et al. (2024); Jiang et al. (2024) also show that LLMs sometimes do not hold consistent personalities.