ExploreSelf: Fostering User-driven Exploration and Reflection on Personal Challenges with Adaptive Guidance by Large Language Models

TL;DR

ExploreSelf addresses the problem of disengagement in writing-based self-reflection by enabling user-driven exploration guided by adaptive prompts from a large language model. The approach combines an initial narrative with theming, Socratic questions, keywords, comments, and AI-generated summaries across three phases, supported by dedicated generative pipelines, to preserve user autonomy while providing tailored guidance. In an exploratory study with 19 Korean adults, participants showed a significant increase in perceived agency and demonstrated diverse patterns of engagement with themes, questions, and summaries, highlighting design implications for AI-guided self-reflection tools. The work contributes a complete design, implementation, and empirical evaluation of an LLM-driven reflective-writing interface and discusses long-term, multi-session considerations, safety, and cultural sensitivity for practical deployment.

Abstract

Expressing stressful experiences in words is proven to improve mental and physical health, but individuals often disengage with writing interventions as they struggle to organize their thoughts and emotions. Reflective prompts have been used to provide direction, and large language models (LLMs) have demonstrated the potential to provide tailored guidance. However, current systems often limit users' flexibility to direct their reflections. We thus present ExploreSelf, an LLM-driven application designed to empower users to control their reflective journey, providing adaptive support through dynamically generated questions. Through an exploratory study with 19 participants, we examine how participants explore and reflect on personal challenges using ExploreSelf. Our findings demonstrate that participants valued the flexible navigation of adaptive guidance to control their reflective journey, leading to deeper engagement and insight. Building on our findings, we discuss the implications of designing LLM-driven tools that facilitate user-driven and effective reflection of personal challenges.

Figures (6)

• Figure 1: The Initial Narrative page of ExploreSelf. Writing the initial narrative ⓐ is a starting point for providing the basis clues to the system. Clicking the 'Start Exploration' button ⓑ leads to the Exploration page (\ref{['fig:interface:components']}-[boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]A).
• Figure 2: Detailed interactions of the Exploration page [boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]A of ExploreSelf. The main scroll panel vertically appends a theme panel ⓐ when the user creates a new theme ⓕ on the Theme selection dialogue [boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]B. The theme panel initially displays only the list of AI-suggested questions regarding the theme [boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]C, and the user can populate the question thread ⓑ by selecting the next question they want to answer ⓘ. While answering a question, users can reveal keywords and request more ⓒ. The system also provides a guide comment ⓓ, which can be regenerated upon request.
• Figure 3: The AI Summary page of ExploreSelf where the users can overview their exploration history ⓐ and the AI-generated summarization ⓑ. The 'View New Summary' button ⓒ generates a new summary. Users can go back to the Exploration page (\ref{['fig:interface:components']}-[boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]A) using the 'Go back' button ⓓ.
• Figure 4: The overview of pipelines to generate themes [boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]A, questions [boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]B, keywords [boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]C, comments [boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]D, and the summary of the exploration history [boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]E in the middle of exploration. Each generative pipeline incorporates an LLM inference that receives the current exploration history starting from the initial narrative as an input. The LLM inference is driven by chain-of-thought style instructions, providing descriptions and rationales for each output to enhance reliability.
• Figure 5: The timelines of participants during the exploration phase. The X-axis indicates elapsed minutes. The yellow bars denote the periods spent writing the initial narratives in the Initial Narrative page (\ref{['fig:system:narrative']}), the cyan bars denote the periods staying in the Exploration page (\ref{['fig:interface:components']}-[boxparam, border-radius=0pt, padding-left=2pt, padding-right=2pt, height=5.5pt, border-width=0pt, background-color=darkgray]A), and the blue bars denote the periods staying in the AI Summary page (\ref{['fig:system:summary']}). Note that participants were able to move between the Exploration and the AI Summary pages.