InnerPond: Fostering Inter-Self Dialogue with a Multi-Agent Approach for Introspection

Paper Title

InnerPond: Fostering Inter-Self Dialogue with a Multi-Agent Approach for Introspection

Publication Info

• Topic area: AI-mediated introspection using multi-agent systems based on Dialogical Self Theory.
• Keywords: Dialogical Self Theory, introspection, multi-agent systems, inner dialogue, large language models, career decision-making, self-reflection, AI-mediated interaction, personal informatics, HCI.

Background and Problem

• Problem / challenge: Existing introspection tools often treat the self as a unified entity, neglecting the plural and dynamic nature of internal perspectives. This limits their ability to support complex decision-making processes involving competing values and aspirations.
• Significance: Addressing this gap is important for fostering deeper self-understanding, particularly in contexts like career decisions, where internal conflicts and negotiations are common.
• Motivation and related work: Prior approaches, including journals, artifacts, and LLM-based systems, have supported introspection but primarily framed users as singular selves. Dialogical Self Theory (DST) offers a framework for understanding the self as a dynamic multiplicity of interacting I-positions, which has not been fully operationalized in interactive systems.

Solution

• Proposed approach: InnerPond, an AI-mediated multi-agent system that externalizes internal perspectives as distinct LLM-based agents (visualized as lotus leaves) to facilitate inter-self dialogue.
• Novelty:
  1. Operationalization of DST through interactive design, enabling users to engage with multiple I-positions.
  2. Introduction of spatial metaphors (lotus pond) to represent coexistence and relational connectedness among inner voices.
  3. Structured scaffolding for introspection across four stages: I-position construction, relational positioning, dialogical exchange, and reflective snapshot.
  4. Empirical insights into inter-self communication and its impact on self-understanding and decision-making.
• Procedure and key techniques:
  • Stage 1: AI-assisted generation and user refinement of I-positions based on personal data.
  • Stage 2: Spatial arrangement of lotus leaves to express relationships among inner voices.
  • Stage 3: Multi-agent dialogue among I-positions, with user mediation and observation.
  • Stage 4: Saving the inner landscape as a temporal self-portrait for reflection and tracking change.

Results

• Concrete findings:
  • Participants retained most AI-generated I-positions (M=11.06), enriched narratives (M=11.82 times), and added new leaves (M=1.67).
  • Dialogues among I-positions revealed unexpected insights and facilitated balanced perspectives on internal conflicts.
  • Visual composition of inner landscapes supported meta-positional reflection, enabling participants to see their selves as interconnected rather than fragmented.
• Advantage over baselines: Unlike traditional introspection tools, InnerPond externalized inner multiplicity, enabling structured dialogue and relational visualization that deepened self-understanding.
• Experiments / evaluation:
  • Participants: 17 young adults (average age = 27.18) deliberating between two career paths.
  • Procedure: Three-week study with pre-survey, in-person session, and follow-up interview.
  • Metrics: Behavioral logs (e.g., I-position modifications, dialogue turns), qualitative interviews, and thematic analysis.
• Limitations and future work:
  • Limited generalizability due to cultural and demographic homogeneity (South Korean university students).
  • Single-session design restricted examination of longitudinal dynamics.
  • Future directions include extending to other domains (e.g., mental health, education) and exploring adaptive scaffolding for diverse user needs.

Summary

InnerPond operationalizes Dialogical Self Theory through an AI-mediated multi-agent system that externalizes inner voices as distinct agents, enabling users to engage in structured inter-self dialogue. Empirical findings demonstrate how externalization, relational visualization, and dialogical exchange support deeper self-understanding, particularly in complex decision-making contexts like career exploration. While tensions between alignment, consistency, and user agency emerged, these dynamics highlight the potential of dialogical introspection systems to scaffold internal negotiations. Future work should address cultural diversity, longitudinal engagement, and domain-specific applications while ensuring ethical safeguards for introspection tools.