Organizations: hci+d lab. Seoul National University Seoul, Republic of Korea · Department of Communication Seoul National University Seoul, Republic of Korea
Introspection is central to identity construction and future planning, yet most digital tools approach the self as a unified entity. In contrast, Dialogical Self Theory (DST) views the self as composed of multiple internal perspectives, such as values, concerns, and aspirations, that can come into tension or dialogue with one another. Building on this view, we designed InnerPond, a research probe in the form of a multi-agent system that represents these internal perspectives as distinct LLM-based agents for introspection. Its design was shaped through iterative explorations of spatial metaphors, interaction scaffolding, and conversational orchestration, culminating in a shared spatial environment for organizing and relating multiple inner perspectives. In a user study with 17 young adults navigating career choices, participants engaged with the probe by co-creating inner voices with AI, composing relational inner landscapes, and orchestrating dialogue as observers and mediators, offering insight into how such systems could support introspection. Overall, this work offers design implications for AI-supported introspection tools that enable exploration of the self's multiplicity.
Figures & tables
Figure 1. InnerPond fosters AI-mediated introspection through multi-agent dialogue. The system transforms users’ I-positions into conversing LLM agents (lotus leaves), supporting dialogue among multiple I-positions within the self. A conceptual visualization of the InnerPond system where users' I-positions become conversing LLM agents represented as lotus leaves on a pond. The image illustrates two example I-positions with contrasting perspectives on career priorities engaging in dialogue. Additional lotus leaves represent other I-positions, demonstrating the system's multi-agent approach to supporting dialogue among multiple I-positions within the self.
Figure 2. Core concepts of Dialogical Self Theory, illustrating the relationship between multiple I-positions in the dialogical space and a meta-position. This figure illustrates the core concepts of Dialogical Self Theory. The internal domain of the self contains multiple I-positions (e.g., Myself, Seeking Recognition,' Myself, Pursuing Stability'), each representing distinct perspectives within the society of mind. The meta-position, depicted as an observer, provides a higher-level vantage point for reflecting on and integrating these diverse internal voices. The dialogical space encompasses both internal and external domains where these I-positions engage in dynamic interactions.
Figure 3. Three-phase metaphor evolution for the dialogical space: From group chat as battle ring (Phase 1), to stones highlighting individuality yet emphasizing separation (Phase 2), to lotus leaves appearing distinct yet sharing roots, enabling a meta-position perspective (Phase 3). Three-phase metaphor evolution showing: a group chat room resembling a battle ring (left), stones arranged on a landscape representing individual I-positions (center), and a pond with lotus leaves sharing roots underwater (right).
Figure 4. I-position extraction and story enrichment pipeline: user knowledge is transformed into distinct I-positions, iteratively refined through scaffolding questions and user responses. This figure illustrates the I-position extraction and story enrichment pipeline. In the extraction stage, user knowledge (career context and profile data) is transformed into distinct I-positions, each composed of a name, viewpoint, and narrative. In the enrichment stage, the system generates scaffolding questions to identify narrative gaps. The user's responses then iteratively refine each I-position's viewpoint and narrative.
Figure 5. The interface for [Stage 1: I-position Construction]: (a) I-positions visualized as lotus leaves on the main pond view, with a profile modal showing name, viewpoint, and narrative; (b) “Story Enrichment” modal with scaffolding questions to refine the narrative; (c) “1:1 Dialogue” modal for direct conversation with a leaf agent. This figure illustrates the Stage 1 (I-position Construction) interface. (a) The main pond view displays I-positions as lotus leaves; clicking a leaf opens a profile modal showing the I-position's name, core viewpoint, and narrative. (b) The Story Enrichment'' modal presents three scaffolding questions to help users refine the I-position's narrative. (c) The 1:1 Dialogue'' modal shows a conversation between the user and a leaf agent, allowing users to explore the I-position's motivations.
Figure 6. The interface for [Stage 2: Relational Positioning]: Users can freely customize their inner landscape by (a) adjusting leaf size, (b) repositioning leaves via drag-and-drop, and (c) selecting colors through the dewdrop-shaped button. This figure illustrates the Stage 2 (Relational Positioning) interface. (a) Users can adjust a leaf's size by hovering over its edge. (b) Leaf positions can be modified via drag-and-drop to arrange the inner landscape. (c) Clicking the dewdrop-shaped button opens a color palette for personal expression. These manipulations allow users to visually compose their inner structure in personally meaningful ways.
Figure 7. The interface for [Stage 3: Dialogical Exchange]: (Left) After selecting two leaf agents, the system suggests three discussion topics based on their relationship; (Right) the group chat where users can participate via “Send” or observe via “Skip.” This figure illustrates the Stage 3 (Dialogical Exchange) interface. (Left) After selecting two leaf agents, the user chooses one of three tailored discussion topics to initiate the conversation. (Right) The group chat shows two leaf agents responding to the selected topic from their distinct perspectives; the user can participate by sending a message or skip to observe the dialogue between agents.
Figure 8. Example of a one-on-one dialogue with [Myself, Diving Deep] (P15). This figure provides an example of a one-on-one dialogue between a user (P15) and their I-position, Myself, Diving Deep.' In the conversation, the user asks for advice on sustaining motivation. The Myself, Diving Deep' persona responds with insights that reflect its core character, suggesting a flexible, curiosity-driven approach over a rigid plan. This exchange demonstrates how users can engage with specific facets of their personality as distinct conversational partners to explore personal challenges and gain new perspectives.
Figure 9. Examples of holistic inner landscapes composed by participants. (a) P4 created a scene of self-compassion, describing one self whispering encouragement to another. (b) P6 arranged the leaves to express gradual personal growth, likened to a musical crescendo. This figure presents two examples of inner landscapes, demonstrating how participants used the size and spatial arrangement of I-position leaves to convey personal narratives. Participant P4 (a) depicted a scene of self-compassion by positioning a large Myself, Believing in Myself' leaf behind Myself, Wanting to Make Meaningful Contribution,' creating a visual impression of one self whispering encouragement to another. Participant P6 (b) arranged numerous leaves in a progressive spread of varying sizes to represent gradual personal growth, likening the composition to a musical crescendo.
Figure 10. Example of a group conversation where P3 pairs [Myself, Yearning for Creative Freedom] and [Myself, Desiring Financial Abundance]. This figure captures a group conversation where a user mediates between two conflicting I-positions: Myself, Yearning for Creative Freedom' and Myself, Desiring Financial Abundance.' The dialogue begins with the user posing a dilemma about balancing creative work with financial stability. Each I-position initially presents its perspective, creating a clear tension. The user then acts as a facilitator, synthesizing their points into a potential compromise. In the final exchange, the I-positions refine this compromise, ultimately agreeing that a stable job can serve as a foundational step towards creative goals, rather than an obstacle. This demonstrates the process of negotiating internal conflicts and reaching a unified, actionable strategy.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Stage
Data Type
Description
Example Metrics
Stage 1: I-Position Construction
I-position profiles
Initial AI-generated and user-modified leaf profiles
Name, core viewpoint, narrative
Profile modifications
User edits to leaf profile
Edit count per leaf
Leaf additions and deletions
User-initiated creation or removal of leaves
Additions ( M=1.67 ), Deletions ( M=1.17 )
1:1 dialogue logs
Chat messages between user and individual leaves
Leaves engaged ( M=5.29 ), Turns ( M=4.92 )
Stage 2: Relational Positioning
Spatial configurations
Position coordinates of each leaf
X, Y coordinates
Visual attributes
Size and color assignments for each leaf
Size value, color
Appendix
Table 1. Overview of System Log Data Collected Across InnerPond Stages
In architectural interior design, miscommunication frequently arises as clients lack design knowledge, while designers struggle to explain complex spatial relationships, leading to delayed timelines and financial losses. Recent advancements in generative layout tools narrow the gap by automating 3D visualizations. However, prevailing methodologies exhibit limitations: rule-based systems implement hard-coded spatial constraints that restrict participatory engagement, while data-driven models rely on extensive training datasets. Recent large language models (LLMs) bridge this gap by enabling intuitive reasoning about spatial relationships through natural language. This research presents an LLM-based, multimodal, multi-agent framework that dynamically converts natural language descriptions and imagery into 3D designs. Specialized agents (Reference, Spatial, Interactive, Grader), operating via prompt guidelines, collaboratively address core challenges: the agent system enables real-time user interaction for iterative spatial refinement, while Retrieval-Augmented Generation (RAG) reduces data dependency without requiring task-specific model training. This framework accurately interprets spatial intent and generates optimized 3D indoor design, improving productivity, and encouraging nondesigner participation. Evaluations across diverse floor plans and user questionnaires demonstrate effectiveness. An independent LLM evaluator consistently rated participatory layouts higher in user intent alignment, aesthetic coherence, functionality, and circulation. Questionnaire results indicated 77% satisfaction and a clear preference over traditional design software. These findings suggest the framework enhances user-centric communication and fosters more inclusive, effective, and resilient design processes. Project page: https://rsigktyper.github.io/AICodesign/
Ren Jian Lim, Rushi Dai
Hong Kong Center for Construction Robotics, The Hong Kong University of Science and Technology
A hallmark of human intelligence is Introspection-the ability to assess and reason about one's own cognitive processes. Introspection has emerged as a promising but contested capability in large language models (LLMs). However, current evaluations often fail to distinguish genuine meta-cognition from the mere application of general world knowledge or text-based self-simulation. In this work, we propose a principled taxonomy that formalizes introspection as the latent computation of specific operators over a model's policy and parameters. To isolate the components of generalized introspection, we present Introspect-Bench, a multifaceted evaluation suite designed for rigorous capability testing. Our results show that frontier models exhibit privileged access to their own policies, outperforming peer models in predicting their own behavior. Furthermore, we provide causal, mechanistic evidence explaining both how LLMs learn to introspect without explicit training, and how the mechanism of introspection emerges via attention diffusion.
Dialogue systems have long supported learner reflections, with theoretically grounded, rule-based designs offering structured scaffolding but often struggling to respond to shifts in engagement. Large Language Models (LLMs), in contrast, can generate context-sensitive responses but are not informed by decades of research on how learning interactions should be structured, raising questions about their alignment with pedagogical theories. This paper presents a hybrid dialogue system that embeds LLM responsiveness within a theory-aligned, rule-based framework to support learner reflections in a culturally responsive robotics summer camp. The rule-based structure grounds dialogue in self-regulated learning theory, while the LLM decides when and how to prompt deeper reflections, responding to evolving conversation context. We analyze themes across dialogues to explore how our hybrid system shaped learner reflections. Our findings indicate that LLM-embedded dialogues supported richer learner reflections on goals and activities, but also introduced challenges due to repetitiveness and misalignment in prompts, reducing engagement.
Paras Sharma, YuePing Sha, Janet Shufor Bih Epse Fofang +6
University of Pittsburgh Pittsburgh, Pennsylvania, USA · Carnegie Mellon University Pittsburgh, Pennsylvania, USA · Bowie State University Bowie, Maryland, USA