Exploring the Design of LLM-based Agent in Enhancing Self-disclosure Among the Older Adults

Research Background and Issues

• What problems or challenges did the authors identify?
  Older adults increasingly face the issue of social isolation, particularly due to shrinking social circles and separation from family members. They lack sufficient opportunities for self-disclosure (sharing personal life, feelings, or thoughts), which is critical for mental health and building intimate relationships. However, due to technological limitations, traditional conversational agents fail to provide deep empathetic feedback, and their overly rigid response patterns cannot meet the more complex emotional needs of older adults.

• Why is this issue important?
  Self-disclosure has been proven to alleviate loneliness, enhance self-esteem, and build psychological resilience. However, older adults often find it difficult to express negative or sensitive topics, such as health issues and financial difficulties, within intimate social circles. Therefore, designing technological systems (e.g., conversational agents based on large language models) that effectively support self-disclosure among older adults can provide significant support in alleviating loneliness and improving their quality of life.

• Research Motivation and Related Work
  Although conversational agents are considered a potential alternative to promote self-disclosure among older adults, most existing studies rely on fixed conversational patterns in predefined scenarios, lacking in-depth exploration of the flexibility and emotional understanding capabilities of large language models (LLMs) in facilitating self-disclosure. Additionally, the impact of different interaction interfaces (e.g., smartphones or robots) on the willingness to self-disclose remains unclear.

Solution

• What methods or solutions did the authors propose?
  This study designed and developed a "Disclosure-Agent" based on large language models, aiming to encourage older adults to engage in self-disclosure on various everyday topics through both robot and smartphone interfaces.

• What are the innovative aspects of this solution?

  1. Utilized the latest large language model (GPT-4) as the foundation, offering dynamic emotional understanding and generation capabilities.
  2. Compared the effects of robot and smartphone interaction interfaces, providing empirical data for future agent design targeting older users.
  3. Developed a modular response framework (e.g., emotional support, clarification, professional advice, encouragement) to generate personalized responses based on the specific needs of older adults.
  4. Proposed design recommendations emphasizing that agents should be "human-like but not human" to gain user trust and promote self-disclosure.

• What are the implementation steps and key technologies used?

  1. System Architecture Design for the Agent:
     • The backend framework relies on multimodal LLMs (e.g., GPT-4), with a visual agent analyzing sensor data (e.g., facial expressions and emotions) and a conversational agent generating personalized feedback.
  2. Hardware Design:
     • The robot interface is based on Raspberry Pi, with cameras and microphones working together to capture users' voice and facial information.
     • The smartphone interface provides similar functionality through an application.
  3. User Experimentation:
     • Recruited 20 older participants, covering 8 daily topics (e.g., health, memory sharing).
     • Collected data on self-disclosure during interactions, such as speech length and emotional expression, and conducted quantitative and qualitative analysis through expert ratings and corpus analysis.
  4. Data Analysis:
     • Quantified self-disclosure levels based on multiple dimensions (e.g., information, thoughts, emotions) and used semantic analysis to explore the impact of topics and interfaces on the willingness to disclose.

Research Findings

• What specific findings were obtained?

  1. Participants were more inclined to engage in self-disclosure through the robot interface, particularly on topics related to health, motivation, and social connections.
  2. Overall, older adults were more willing to share with robots and perceived them as offering higher privacy security compared to smartphones.
  3. Data showed that while participants subjectively believed robots were more suitable for self-disclosure, there was no significant difference in the level of detail in disclosed content between the two interfaces.
  4. Summarized three major needs of older adults in self-disclosure: obtaining information, seeking resonance, and functional assistance.

• What advantages does it have compared to existing solutions?

  1. Enhanced the intelligence level of conversational agents, moving beyond fixed topic ranges to dynamically understand and respond to more complex emotional and practical needs.
  2. Provided a comparative study of robot and smartphone interfaces, offering key practical data for future technology design for older adults.
  3. Modularized content generation, enabling the agent to adapt more easily to different contexts and user needs.

• What were the experimental or evaluation results?

  1. By analyzing linguistic variables across various topics (e.g., frequency of first-person pronouns and text length), it was found that participants tended to express more content on "memory" topics, while being relatively reserved on "financial" topics.
  2. Topics related to social connection and sharing joy were more likely to elicit thoughts and emotional expression from older adults.
  3. Experimental data also indicated that participants expected the agent to not only provide emotional support but also assist with specific tasks (e.g., making medical appointments).

• Limitations and Future Directions

  1. Sample Limitations: Participants were all from the same city in China, and cultural and regional biases may limit the generalizability of the results.
  2. Age Group Bias: The study focused more on relatively younger older adults, without in-depth exploration of the 85+ age group.
  3. Future Directions:
     • Expand the sample range to test applicability across diverse cultures and age groups.
     • Explore how agents can improve older adults' social connections over the long term, rather than merely serving as a substitute solution.
     • Further refine ethical design to prevent over-reliance on agents or potential privacy violations.

Through this study, the authors not only demonstrated the potential application of LLM-based Disclosure-Agents in promoting self-disclosure among older adults but also highlighted important directions for optimizing their design and ethical practices. This lays a foundation for future developments in HCI and assistive technologies for older adults.