The Interaction Layer: An Exploration for Co-Designing User-LLM Interactions in Parental Wellbeing Support Systems

Research Background and Issues

• What problems or challenges did the authors identify?
  Many new parents face significant emotional and physical stress, while traditional psychological and parenting support is often limited by time, cost, or accessibility. Although AI technology has the potential to improve accessibility and reduce the stigma surrounding mental health, users often reject these systems due to issues with explainability and reliability. Specifically, in the context of parenting support, users desire interactions that are more natural, controllable, and capable of reflecting contextual and personalized needs.

• Why is this issue important?
  During the parenting phase, especially the postpartum period, parents require effective and reliable support systems to alleviate stress, enhance physical and mental well-being, and foster the development of parent-child relationships. Exploring AI technologies that meet these needs can not only address current service gaps but also have broad practical applications and social impact.

• Research motivation and related work
  The authors aim to explore the integration of large language models (LLMs) and co-design methods in the parenting support domain to improve user experience. By investigating the technical shortcomings of existing parenting support systems, the study seeks new methods to enhance the quality of AI-parent interactions.

Solution

• What methods or solutions did the authors propose?
  The authors designed a large language model-driven system called "NurtureBot," which improves its interaction layer through co-design engagement with users. The core design is based on three principles: helping users understand the system, controlling interactions, and continuously improving interaction outcomes. The research process is divided into four parts: prototype testing, user needs co-design, interaction layer improvement, and final version validation.

• What are the innovative aspects of the solution?
  The solution's innovations include:

  • Proposing an interaction layer architecture that enables the system to adjust its behavior more effectively through dynamic state transitions and user prompts.
  • Applying user co-design techniques to directly integrate user-generated suggestions and conversational examples (few-shot prompting) into the system, enhancing interaction naturalness.
  • Introducing the "Understand, Control, and Improve" model, enabling the AI to better adapt proactively to user needs.

• What are the implementation steps? What key technologies were used?
  The methods include the following implementation steps:

  1. Building the initial prototype (NurtureBot v1): Designing a basic AI conversational model using zero-shot prompting to provide empathetic dialogue, health-related exercises, and parenting information.
  2. User needs collection and co-design (ARC method): Participants helped prioritize issues and role-played as NurtureBot to generate improved conversational suggestions.
  3. Introducing the interaction layer (NurtureBot v2): Integrating user-designed conversational samples into the system's prompt architecture while enhancing the dynamic adaptability of interactions.
  4. Improving the final version (NurtureBot v3): Addressing subsequent issues, such as shortening response lengths, increasing contextual actionability in prompts, and adding multi-directional validation mechanisms.

Research Outcomes

• What specific outcomes were achieved?
  Through iterative design and testing, the authors significantly improved the user experience of the chatbot system. The final version, NurtureBot v3, achieved a CUQ score of 91.3, the highest reported in the field.

• What advantages does it have compared to existing solutions?
  The system's advantages include:

  • Enhanced naturalness in AI-parent interactions, significantly reducing awkward or failed conversations.
  • Real-time conversational optimization integrated with user-participatory design, making it more aligned with user needs.
  • An effective interaction architecture model capable of flexibly transitioning between different task states while maintaining high transparency and user control.

• What were the experimental or evaluation results?
  Results from the four-stage study indicated that the co-design improvements in subsequent versions of NurtureBot led to enhancements in overall usability, interaction comprehension, and control capabilities. Participant satisfaction with the final version increased significantly.

• Limitations and future directions
  Limitations include:

  • The test group was primarily based in the UK, making it difficult to generalize results to regions with different socio-cultural contexts.
  • The absence of longitudinal usage data limits the evaluation of the system's impact on long-term user relationships.
  • The system has not yet addressed risks such as user over-reliance on the machine or misinformation.

  Suggested future directions:

  • Further personalization and localization of the system, incorporating memory functions and multimodal interactions.
  • Strengthening connections to community resources to enhance contextual relevance.
  • Exploring ethical issues, such as mitigating risks associated with users overly anthropomorphizing interaction content.

Conclusion

By combining user co-design principles with advanced LLM technology, the research provides an innovative solution for parenting support systems, setting a new benchmark for designing digital health technologies. Through continuous iteration and experimental validation, the study demonstrates the feasibility of AI interaction systems that can effectively improve user experience, offering valuable design frameworks and insights for the development of similar applications in the future.