# Character-LLM: A Trainable Agent for Role-Playing

**Authors:** Yunfan Shao, Linyang Li, Junqi Dai, Xipeng Qiu  
**arXiv:** [2310.10158](https://arxiv.org/abs/2310.10158)  
**Date:** 2023-10-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.10158v2/

## Key points

- Authors explore potential of Large Language Models (LLMs) as agents simulating human behaviors
- Highlight remarkable ability of LLMs to comprehend human instructions and generate high-quality texts
- Propose training an agent with the profile, experiences, and emotional states of a specific person instead of relying on limited prompts
- Introduce Character-LLM to teach LLMs to embody specific historical figures such as Beethoven, Queen Cleopatra, Julius Caesar, among others
- Aim to evaluate efficacy through a test playground where trained agents are interviewed
- Experimental results offer insights into building future simulacra representing various aspects of humankind
- Research contributes to advancing capabilities of language models in simulating complex human behaviors and personalities through targeted training approaches

## Summary

In their paper titled "Character-LLM: A Trainable Agent for Role-Playing," authors Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu explore the potential of Large Language Models (LLMs) to act as agents simulating human behaviors. They highlight the remarkable ability of LLMs to comprehend human instructions and generate high-quality texts. This prompts them to investigate whether these models can go beyond simple human behaviors and simulate individuals in a more complex manner. The authors propose training an agent with the profile, experiences, and emotional states of a specific person instead of relying on limited prompts to direct ChatGPT API. Their approach involves introducing Character-LLM, which teaches LLMs to embody specific historical figures such as Beethoven, Queen Cleopatra, Julius Caesar, among others. By modifying profiles to reflect the experiences of these characters and training models to serve as personal simulacra based on these experiences, they aim to evaluate the efficacy of their methodology through a test playground where trained agents are interviewed. The experimental results yield intriguing observations that offer insights into building future simulacra representing various aspects of humankind. This innovative research contributes to advancing the capabilities of language models in simulating complex human behaviors and personalities through targeted training approaches. The authors' work opens up new possibilities for leveraging LLMs in role-playing scenarios and creating more nuanced simulations of historical or fictional characters.