CloChat: Understanding How People Customize, Interact, and Experience Personas in Large Language Models

Title of the Paper

CloChat: Understanding How People Customize, Interact, and Experience Personas in Large Language Models

Paper Information

• Domain: User interaction and personalization design in large language models (LLMs)
• Keywords: Personalization, LLMs, interactive systems, user customization, conversational assistants in large language models

Research Background and Problem

• Challenges Identified by Authors:

  • Current LLM-based conversational agents often exhibit a uniform, fixed personality, which fails to meet the diverse needs of users.
  • While users can customize agents through prompt engineering, the process is complex and unintuitive, limiting accessibility for general users.
  • Although prior research suggests that persona-based design leads to more satisfying and human-like conversational experiences, there is limited exploration of how users construct and interact with persona designs in LLMs.

• Significance:

  • Persona customization can significantly enhance user trust, interaction quality, and immersion, driving the design of more personalized interactive systems.
  • Investigating the long-term relationships between users and customized agents can improve LLM adaptability and sustained interaction in practical applications.

• Related Work:

  • LLMs have achieved remarkable progress in text generation and contextual understanding but still face significant shortcomings in personalization.
  • Existing studies have explored agents with fixed roles and domain-specific characteristics but fail to address users' specific emotional and task-oriented needs.

Solution

• Proposed Approach:

  • Develop CloChat, an interface that allows users to easily customize personalized agents and interact with them. CloChat includes:
    • CloChat Design Lab: Enables users to define agent demographic information, personality traits, nonverbal expressions (e.g., emojis), knowledge domains, and visual appearances.
    • CloChat Chatroom: Allows users to experience conversational interactions with customized agents.

• Innovations:

  • Offers a highly intuitive form-based interface, completely simplifying the complex "prompt engineering" process.
  • Supports visual customization by generating agents' personal images using DALL-E2, enhancing the personalization experience.
  • Integrates an experimental "preview feature" that allows users to perceive agent behaviors in real-time during adjustments.

• Implementation Steps and Techniques:

  • Converts all user inputs into JSON format and translates them into natural language expressions via GPT-4.
  • Utilizes GPT-4 for generating and responding as conversational agents.
  • Simplifies the customization process through tutorials and lab-style interfaces, while providing multiple visual options.

Research Outcomes

• Specific Results:

  • Users who customized agents with CloChat reported significantly increased trust and emotional connection with the system.
  • Conversations with CloChat demonstrated greater diversity compared to standard ChatGPT, especially in interactions across different scenarios.
  • Personalized agent dialogues elicited deeper emotional engagement from users, extending beyond task-oriented interactions.

• Comparison with Existing Solutions:

  • Compared to ChatGPT, CloChat proved more effective and engaging, better sustaining user interest and willingness for continued interaction.
  • Customizable visual features and personalization options further optimized user experience, which are absent in ChatGPT.

• Limitations and Future Directions:

  • The study did not cover broader cultural and linguistic experiences, nor did it explore long-term user-agent interaction relationships.
  • The complexity of personalization may pose a barrier for first-time users.
  • Ethical concerns, such as designing agents that mimic real-world individuals, require further investigation.
  • Future research should explore automated personality learning and more flexible customization methods to address current limitations.

Design Insights

• Design Directions:
  1. Provide user-friendly customization options and simplify the initial setup process.
  2. Prioritize the development of dynamic agents with strong contextual adaptability to meet diverse user needs.
  3. Advocate for strict ethical guidelines to limit designs that mimic real individuals or pose privacy risks.
  4. Explore automated personality learning algorithms to match user preferences and contexts, optimizing interaction experiences.