Character-LLM: A Trainable Agent for Role-Playing (2310.10158v2)

Abstract: LLMs can be used to serve as agents to simulate human behaviors, given the powerful ability to understand human instructions and provide high-quality generated texts. Such ability stimulates us to wonder whether LLMs can simulate a person in a higher form than simple human behaviors. Therefore, we aim to train an agent with the profile, experience, and emotional states of a specific person instead of using limited prompts to instruct ChatGPT API. In this work, we introduce Character-LLM that teach LLMs to act as specific people such as Beethoven, Queen Cleopatra, Julius Caesar, etc. Our method focuses on editing profiles as experiences of a certain character and training models to be personal simulacra with these experiences. To assess the effectiveness of our approach, we build a test playground that interviews trained agents and evaluates whether the agents \textit{memorize} their characters and experiences. Experimental results show interesting observations that help build future simulacra of humankind.

• The paper introduces Character-LLM as a method to fine-tune language models for simulating human characters with consistent personas.
• It employs experience reconstruction, targeted data upload, and protective experiences to maintain character authenticity.
• Experimental results demonstrate improved character consistency and reduced hallucination compared to baseline models like Alpaca and Vicuna.

Character-LLM: A Trainable Agent for Role-Playing

The paper "Character-LLM: A Trainable Agent for Role-Playing" explores an intriguing approach to simulating human characters using LLMs. Unlike typical prompting methods, the authors propose training specific models to embody historical or fictional figures, allowing these models to maintain consistent personas across varied interactions.

Methodology Overview

The paper introduces Character-LLM, a method in which LLMs are fine-tuned to simulate specific characters such as Beethoven or Cleopatra. The approach consists of three main components:

1. Experience Reconstruction: The authors compile character profiles and use them to create detailed scenes that mimic the experiences and interactions of the characters. This process leverages the capabilities of models like LLaMA to simulate the intricate traits and personal histories of these figures.
2. Experience Upload: LLMs are trained using the reconstructed experiences, effectively loading these narratives into the model, thus allowing it to simulate the character with greater authenticity.
3. Protective Experiences: This involves training models to forget or avoid knowledge that contradicts the character's historical or fictional context, preventing anachronistic or irrelevant responses.

Experimental Results

The evaluation focuses on model performances across several dimensions: memorization, values, personality, hallucination avoidance, and stability. The trained agents show consistent character portrayal, outperforming baseline models like Alpaca and Vicuna. Specifically, the results demonstrate that Character-LLMs excel in maintaining the distinct traits and knowledge scopes of the characters they simulate. Notably, the introduction of protective experiences helps mitigate hallucination, restricting knowledge to what is appropriate for the simulated persona.

Implications and Future Directions

The implications of this research are manifold, especially for applications requiring role-based interactions, such as NPCs in games, educational tools, and historical studies. Trainable agents could become instrumental in fields like social sciences by providing insights into historical and theoretical character interactions.

The paper suggests potential developments in integrating multimodal data, increasing the vividness and diversity of character experiences beyond textual content. Future work could explore larger model architectures or more extensive pre-training data to enhance the authenticity and complexity of the role-play simulations.

Conclusion

Character-LLM represents a significant step in specialization of LLMs for simulating human-like characters. This approach underscores the potential of trainable agents to augment human-computer interaction by providing consistent, believable personas, while also addressing the challenges inherent in retaining the nuances of individual character traits. The research opens new avenues for the use of AI in immersive simulations and interactive storytelling.