ExploreSelf: Interactive Reflection

• ExploreSelf is an LLM-driven interactive reflection system that generates adaptive guidance based on users' narratives.
• It employs chain-of-thought prompts to produce personalized themes, Socratic questions, keywords, and AI summaries that guide reflective writing.
• Empirical studies show the tool enhances user agency, though occasional misalignment in guidance indicates a need for real-time adaptive refinements.

ExploreSelf is an LLM-driven interactive system designed to facilitate user-driven exploration and reflection on personal challenges. The system dynamically generates adaptive guidance—such as questions, themes, and comments—using LLMs, with the aim of empowering individuals to flexibly organize, reflect upon, and gain insights into their experiences. ExploreSelf departs from prescriptive or linear writing interventions by providing users with a high degree of control over their reflective process, enabling the depth and direction of exploration to be tailored on demand.

1. System Architecture and Interactive Flow

ExploreSelf’s architecture is organized around three sequential, user-centered phases:

1. Initial Narrative: Users freely compose a narrative about a personal challenge, unconstrained by predetermined prompts.
2. Exploration Phase: The system analyzes the narrative to extract salient “themes” and presents these to the user, who selects a theme for deeper inquiry. For each chosen theme, the system generates a thread of Socratic, open-ended questions designed to stimulate reflection and self-exploration.
3. AI Summary Phase: At any point, users can access a synthesized AI-generated summary that distills their evolving reflections, including themes, responses, and key insights.

The interface supports high interactivity and flexibility. Users can toggle AI-generated keywords (providing cognitive scaffolding) and tailored comments (tips, sub-questions, or affirmations) on or off, allowing for granular, on-demand guidance. Navigation between question threads, themes, and summary views is non-linear, empowering users to control their trajectory and revisit or skip elements as desired.

2. LLM-Driven Adaptive Guidance Mechanisms

All core guidance elements in ExploreSelf are generated by LLMs using carefully designed prompt pipelines grounded in chain-of-thought principles and a “therapeutic assistant” persona. These generative routines encompass:

• Theme Generation: The user’s narrative is processed by the LLM to extract primary themes, alternative phrasings, and salient quotes, through prompts using XML-structured instructions and chain-of-thought to ensure themes align with user language and tone.
• Question Generation: When a user selects a theme, the LLM synthesizes multiple Socratic questions, prompting users to explore the theme through different perspectives or levels of depth.
• Keyword and Comment Generation: Keywords highlight concepts for further consideration, while comments offer meta-cognitive support or gentle redirection.
• Summary Generation: The entire exploration session—comprising themes, user responses, and AI content—is condensed into a narrative summary, reflecting the progression of thoughts and insights.

Technically, the backend is realized with TypeScript and Express.js; the frontend uses React.js; data is stored in MongoDB. LLM interactions are orchestrated with LangChain.js, using the OpenAI GPT-4 ChatCompletion API (in particular, the gpt-4o variant for improved Korean language support).

3. User Experience, Agency, and Engagement

An exploratory paper with 19 participants (session durations of 30–60 minutes) demonstrated that ExploreSelf enabled diverse patterns of reflective engagement. Users could either concentrate on a single theme with multiple rounds of guidance or navigate across various themes and prompts opportunistically. The system’s non-directive, flexible structure supported a significant increase in users’ perceived agency, as measured by the Pathways subscale of self-determination (average increase ≈2.63 points).

Qualitative feedback indicated that adaptive guidance sustained focus, introduced novel perspectives, and helped users organize diffuse thoughts. However, occasional mismatches between AI-generated prompts and user intent could momentarily disrupt the reflective process. Overall, users reported that the ability to modulate the depth and breadth of exploration contributed to deeper insight and a stronger sense of control over personal reflection.

4. Design Implications and Guidelines for LLM-Driven Reflection Tools

Findings from ExploreSelf highlight critical design considerations for future LLM-driven reflective systems:

• Balance of Guidance and Autonomy: Systems should provide scaffolding without overtly constraining user direction. Multiple adaptive elements (themes, questions, keywords, comments) support a spectrum of user preferences.
• Iterative and Summative Support: The AI-generated summary aids users in contextualizing the evolution of their thought processes, suggesting the value of session memory and long-term summarization features.
• Cultural and Contextual Calibration: Effective reflection tools must accommodate cultural and linguistic diversity, necessitating LLM prompt tuning and appropriate persona design for different user groups.
• Dynamic Adjustment to User Focus: Interfaces should allow easy correction or bypassing of AI suggestions that are misaligned with user intent, maintaining high user agency and engagement.

5. Technical Specification

ExploreSelf’s implementation integrates resilient LLM prompt pipelines and stateful interaction management. Key mechanisms include:

• Prompt Engineering: Chain-of-thought prompts and XML-structured inputs for theme generation; explicit Socratic role instructions in question and comment generation.
• Frontend/Backend Communication: Real-time toggling and presentation of AI-generated elements via AJAX and state management in React.
• Data Management and Privacy: Session logs and user-generated content are stored securely in MongoDB. No claims regarding privacy-preserving guarantees are stated.

The following table summarizes the mapping between user actions and LLM-guided responses:

User Action	LLM-Generated Guidance	Interaction Outcome
Write initial narrative	Themes, keywords, AI summary	Launches exploration phase
Select a theme	Socratic question thread	Enables deep probe on chosen topic
Respond to AI question	More targeted questions/comments	Fosters iterative reflection
Toggle keywords or comments	Contextual scaffolds	Adjustable guidance granularity
Request AI summary	Narrative condensation	High-level reflection/closure

6. Empirical Results and Limitations

The participant paper validated that adaptive LLM guidance can enhance user agency and engagement in reflective writing. However, misalignments between system-generated prompts and user focus can occur, highlighting the need for finer-grained feedback and interactive override mechanisms. These findings suggest avenues for further work on incorporating user feedback for real-time adaptation, longitudinal support (multi-session memory), and deeper personalization.

7. Prospective Extensions and Research Directions

Research into systems like ExploreSelf points toward:

• Leveraging LLMs not only for guidance but as memory units that recognize recurring user themes across sessions.
• Further refining the balance of prompt-based scaffolding versus user control, possibly through reinforcement learning from implicit user feedback.
• Expanding support for multilingual, multicultural, and domain-specific reflection, potentially by fine-tuning LLMs on contextually rich, localized datasets.
• Exploring ethical considerations and robust privacy frameworks for storing and processing deeply personal reflective content.

In summary, ExploreSelf is an LLM-driven reflective writing tool that operationalizes adaptive, user-driven exploration by dynamically generating themes, Socratic questions, scaffolding keywords, and AI summaries, providing a flexible environment that empowers users to engage in deeply personalized and organized self-reflection (Song et al., 15 Sep 2024).