This survey explores the growing field of role-playing with language models, tracing their evolution from early persona-based models to advanced character-driven simulations enabled by Large Language Models (LLMs). Initially limited to simple persona consistency, role-playing tasks now involve complex character portrayals, including consistency, behavioral alignment, and character attractiveness. The survey provides a comprehensive taxonomy of key components in designing these systems, including data, models and alignment, agent architecture, and evaluation. It outlines current methodologies, challenges, and future research directions to improve the depth and realism of role-playing applications. The goal is to guide future research by offering a structured overview of current methodologies and identifying potential areas for improvement. Related resources and papers are available at https://github.com/nuochenpku/Awesome-Role-Play-Papers. Role-playing with language models differs from generic assistants, as the primary expectation is alignment with specific personas or characters rather than helpfulness. This introduces a dynamic that can sometimes contradict traditional notions of helpfulness. For instance, being helpful may become contradictory when the role is that of an adversary. The evolution of sequence-to-sequence learning has enabled neural networks to generate dialogue responses consistent with both context and persona. Early models like Zhang et al. (2018) used generative profile memory networks to generate personal responses, laying the groundwork for future role-playing works. Subsequent advancements with models like BERT brought significant changes to the use of language models for role-playing, despite inherent limitations. In the era of LLMs, a paradigm shift occurred, expanding the scope of role-playing tasks far beyond simple persona adherence. Current research no longer confines itself to rigid personas but explores nuanced aspects of role enactment, such as character consistency, behavioral alignment, and overall attractiveness of the character portrayal. These dimensions aim to create more immersive and believable character simulations that maintain continuity over interactions and adapt dynamically to dialogue contexts. The progress of LLM-based role-playing has led to rapid expansion in academic research and practical applications, exemplified by platforms like Character AI, Xingye, and Maopaoya. Despite the promising potential of role-playing with language models, research in this domain remains in its early stages, marked by complexities and challenges. The goal of this survey is to understand the crucial mechanisms and methodologies that enable role-playing through text-based interactions. To achieve a thorough understanding, we introduce a detailed taxonomy to systematically examine the critical components involved in designing role-playing language models. The proposed taxonomy includes: Data, Models & Alignment, Agent Architecture, and Evaluation. This framework aims to not only detail how role-playing functions within these systems but also to highlight how it can be optimized and evaluated for a variety of applications.This survey explores the growing field of role-playing with language models, tracing their evolution from early persona-based models to advanced character-driven simulations enabled by Large Language Models (LLMs). Initially limited to simple persona consistency, role-playing tasks now involve complex character portrayals, including consistency, behavioral alignment, and character attractiveness. The survey provides a comprehensive taxonomy of key components in designing these systems, including data, models and alignment, agent architecture, and evaluation. It outlines current methodologies, challenges, and future research directions to improve the depth and realism of role-playing applications. The goal is to guide future research by offering a structured overview of current methodologies and identifying potential areas for improvement. Related resources and papers are available at https://github.com/nuochenpku/Awesome-Role-Play-Papers. Role-playing with language models differs from generic assistants, as the primary expectation is alignment with specific personas or characters rather than helpfulness. This introduces a dynamic that can sometimes contradict traditional notions of helpfulness. For instance, being helpful may become contradictory when the role is that of an adversary. The evolution of sequence-to-sequence learning has enabled neural networks to generate dialogue responses consistent with both context and persona. Early models like Zhang et al. (2018) used generative profile memory networks to generate personal responses, laying the groundwork for future role-playing works. Subsequent advancements with models like BERT brought significant changes to the use of language models for role-playing, despite inherent limitations. In the era of LLMs, a paradigm shift occurred, expanding the scope of role-playing tasks far beyond simple persona adherence. Current research no longer confines itself to rigid personas but explores nuanced aspects of role enactment, such as character consistency, behavioral alignment, and overall attractiveness of the character portrayal. These dimensions aim to create more immersive and believable character simulations that maintain continuity over interactions and adapt dynamically to dialogue contexts. The progress of LLM-based role-playing has led to rapid expansion in academic research and practical applications, exemplified by platforms like Character AI, Xingye, and Maopaoya. Despite the promising potential of role-playing with language models, research in this domain remains in its early stages, marked by complexities and challenges. The goal of this survey is to understand the crucial mechanisms and methodologies that enable role-playing through text-based interactions. To achieve a thorough understanding, we introduce a detailed taxonomy to systematically examine the critical components involved in designing role-playing language models. The proposed taxonomy includes: Data, Models & Alignment, Agent Architecture, and Evaluation. This framework aims to not only detail how role-playing functions within these systems but also to highlight how it can be optimized and evaluated for a variety of applications.