Title: 2604.10545 ArXiv: 2604.10545 Enhanced Self-Learning with Epistemologically-Informed LLM Dialogue Title: Content selection saved. Describe the issue below: Description: License: CC BY 4.0 arXiv:2604.10545v1 [cs.HC] 12 Apr 2026 Enhanced Self-Learning with Epistemologically-Informed LLM Dialogue Yi-Fan Cao 0000-0002-5892-5052 Hong Kong University of Science and Technology Hong Kong China ycaoaw@connect.ust.hk , Kento Shigyo 0000-0002-5095-7500 Hong Kong University of Science and Technology Hong Kong China kshigyo@connect.ust.hk , Yitong Gu 0009-0001-2890-5448 Hong Kong Baptist University Hong Kong China yitonggu@life.hkbu.edu.hk , Xiyuan Wang 0009-0008-1839-2010 ShanghaiTech University Shanghai China wangxy7@shanghaitech.edu.cn , Weijia Liu 0009-0001-6332-5705 Hong Kong University of Science and Technology (Guangzhou) Guangzhou China wliu383@connect.hkust-gz.edu.cn , Yang Wang The University of Hong Kong Hong Kong China yang.wang@hku.hk 0000-0002-8903-2388 , David Gotz 0000-0002-6424-7374 University of North Carolina at Chapel Hill Chapel Hill North Carolina USA gotz@unc.edu , Zhilan Zhou 0000-0003-1236-1287 University of North Carolina at Chapel Hill Chapel Hill North Carolina USA zzl@cs.unc.edu and Huamin Qu Hong Kong University of Science and Technology Hong Kong China huamin@ust.hk 0000-0002-3344-9694 Abstract. Large Language Models (LLMs) have advanced self-learning tools, enabling more personalized interactions. However, learners struggle to engage in meaningful dialogue and process complex information. To alleviate this, we incorporate epistemological frameworks within an LLM-based approach to self-learning, reducing the cognitive load on learners and fostering deeper engagement and holistic understanding. Through a formative study (N=26), we identified epistemological differences in self-learner interaction patterns. Building upon these findings, we present CausaDisco , a dialogue-based interactive system that integrates Aristotle’s Four Causes framework into LLM prompts to enhance cognitive support for self-learning. This approach guides learners’ self-learning journeys by automatically generating coherent and contextually appropriate follow-up questions. A controlled study (N=36) demonstrated that, compared to baseline, CausaDisco fostered more engaging interactions, inspired sophisticated exploration, and facilitated multifaceted perspectives. This research contributes to HCI by expanding the understanding of LLMs as educational agents and providing design implications for this emerging class of tools. Large Language Models (LLMs), Sensemaking, Human-AI Interaction, Epistemology, Self-Learning † † copyright: acmlicensed † † ccs: Human-centered computing Interactive systems and tools † † ccs: Applied computing Education Figure 1. CausaDisco provides users with original learning materials (A) and a core concept graph (B) to facilitate self-learning. Once a user initiates a dialogue, the LLM chatbot offers preliminary answers and then automatically generates four epistemologically-informed follow-up questions (C) to encourage deeper exploration. Concurrently, CausaDisco creates a query tree map (D) to assist users in managing their conversation records. When users are satisfied with the results of multi-turn dialogues (T1-3), they can proceed to explore and learn about new topics. This figure presents the interface and interaction design of CausaDisco, which enhances self-learning by providing original materials and a core concept graph for guidance. When users start a dialogue, the LLM chatbot offers initial answers and generates four follow-up questions to promote deeper understanding. Simultaneously, a query tree map is created to help users manage conversation records. Once users are satisfied with multi-turn dialogues, they can explore new topics. 1. Introduction The rise of Large Language Models (LLMs) has revolutionized education, introducing a new generation of dialogue-driven learning tools (Luo et al. , 2022 ; Chen et al. , 2023b ; Zhang et al. , 2023 ) . These tools have greatly improved the efficiency of independent learning, supporting a wide range of tasks, from spoken language practice (Huang et al. , 2022 ) and text editing (Kim, 2023 ) to literature searches (Zheng et al. , 2024a , b ) and information summarization (Reddy and Guha, 2023 ) . With the adoption of prompt engineering techniques, the potential of LLM-based educational tools to facilitate deeper learning and knowledge construction has been further explored (Suh et al. , 2023 ) . These dialogue-driven tools can help streamline the cognitive processes involved in sensemaking, enabling learners to acquire, comprehend, integrate, and apply new knowledge more effectively (Abdelghani et al. , 2024 ; Gero et al. , 2024 ) . Although LLM-based educational tools promote a flexible and intuitive “search-as-learning” sensemaking process (Liu et al. , 2024a ) , many still require substantial effort from learners to engage in meaningful dialogue and process complex information. This is particularly challenging when using less customized LLM tools, as learners often spend considerable time crafting precise questions to elicit responses aligned with their specific needs. Such exploratory and iterative dialogue-driven interactions can be especially demanding for less proactive learners accustomed to more passive input-driven sensemaking. Consequently, even with LLM-based tools, many learners still encounter substantial, yet often hidden, cognitive load. Extensive studies have focused on enhancing the contextuality (Akhtar et al. , 2019 ; Shihab et al. , 2023 ; Chen et al. , 2023a ) and workflow (Suh et al. , 2023 ; Gao et al. , 2024 ; Shankar et al. , 2024 ) of LLM-based educational tools to support learners’ sensemaking processes. However, existing work often centers on domain-specific applications (Sheng et al. , 2023 ; Chen et al. , 2024b ) or caters to early adopters already proficient in strategically engaging with LLM chatbots (Zheng et al. , 2024b ) . This focus overlooks a fundamental aspect of human-AI interaction: sensemaking is a highly individualized cognitive process. This process is shaped by one’s epistemological schema , which represents structured beliefs about the nature , source , and justification of knowledge (Hofer and Pintrich, 1997 ; Brownlee et al. , 2002 ; Bråten, 2010 ; Baker and Anderman, 2020 ) . Consequently, a critical gap exists in understanding how diverse mental models, reflected in different LLM interaction behaviors influence the design of effective, usable, and inclusive educational tools. To bridge this knowledge gap, we conducted a formative study with 26 participants. Data collection was triangulated (Carter, 2014 ) through dialogue records captured by a probe system, online surveys, and semi-structured interviews 1 1 1 Please refer to the supplementary materials for details of our survey and semi-structured interview protocols. . Through a thematic analysis of the collected data, we derived a query taxonomy that revealed distinct interaction behaviors, while also examining participants’ underlying epistemological schemas. The synthesis of these analyses identified four interaction patterns : proactive , validation-seeking , content-focused , and receptive . We revealed key differences by analyzing these patterns within the epistemological framework of knowledge nature, source, and justification, particularly between proactive and receptive patterns. The proactive pattern was characterized by active knowledge construction, evidenced by participants formulating effective questions, requesting elaborations, and synthesizing information. Conversely, the receptive pattern was characterized by fewer questions, less exploration, and a preference for receiving direct answers from LLMs rather than actively building understanding. Furthermore, the proactive pattern was associated with high-performing participants, while the receptive pattern was often observed in those who struggled in the learning process. The observed epistemological differences and learning outcomes across different patterns highlight the need for tools to accommodate diverse learner needs. We identified three primary challenges in designing inclusive learner-LLM interactions : limited interactivity , inefficient information verification , and confirmation bias . Addressing these challenges requires educational tools that support: (1) lower cognitive load, (2) sustained engagement, (3) multidimensional thinking, (4) metacognition, and (5) validation assistance. Guided by these requirements, we developed CausaDisco , a proof-of-concept system designed to support more effective learner-LLM interactions. CausaDisco comprises four main views (see Fig. 1 ): an Embedded Content View providing access to original learning materials; a Concept Graph visualizing relationships between core concepts; a Q&A Conversation View prompting epistemologically-informed follow-up questions; and a Tree Map View displaying the underlying query logic. Specifically, in the Q&A Conversation View , we incorporate epistemological frameworks with LLM prompt engineering to enhance self-learners sensemaking. This approach is inspired by pedagogical research demonstrating the benefits of structured, epistemologically-informed frameworks for critical thinking, analytical depth, and knowledge acquisition in self-learning contexts (Peters, 2000 ; Bråten, 2010 ; Krasmann, 2020 ) . Drawing on these insights, as well as the interaction patterns distilled from our formative study, we identified the need for a systematic approach to generating follow-up questions that support comprehensive sensemaking. We selected Aristotle’s Four Causes (Hocutt, 1974 ; Falcon, 2006 ) –a classical epistemological framework for knowledge construction–due to its remarkable alignment with our empirically derived query taxonomy. This framework prompts users to consider the Material , Formal , Efficient , and Final causes of concepts, effectively mapping to the diverse types of questions learners naturally ask. To evaluate the efficacy of CausaDisco , we conducted a controlled user study with 36 participants employing a between-subjects design. We assessed both objective and subjective measures, including quiz scores and Likert-scale ratings to assess engagement, efficiency, comprehensiveness, sophistication, and usability of the system. Our results indicate that CausaDisco significantly facilitates users’ sensemaking process, fostering a more engaging learning experience compared to the baseline condition. In summary, our core contributions are as follows: • A taxonomy of interaction patterns in self-learning with LLMs, identifying key epistemological differences and outlining design requirements for supporting user sensemaking. • CausaDisco , an interactive proof-of-concept system for self-learning. The system leverages Aristotle’s Four Causes theoretical framework for LLM prompt engineering to guide user sensemaking and includes features for knowledge synthesis and learning journey tracking. • Empirically-grounded design implications for dialogue-driven LLM educational tools, derived from a within-subjects evaluation of CausaDisco ’s efficacy in enhancing interactivity and sensemaking. 2. Related Work Given the interdisciplinary nature of this research, our literature review integrates technical, epistemological, and pedagogical perspectives. It is structured as follows: 1) LLM-based educational agents ; 2) Integrating epistemology for LLM sensemaking ; and 3) Toward effective learner-LLM interactions . 2.1. LLM-Based Educational Agents LLM-based educational agents advance technology-enhanced learning (TEL) by creating personalized learning experiences (Gan et al. , 2023 ; Lin, 2023 ; Huber et al. , 2024 ) . These agents improve intelligent tutoring systems (ITS) through real-time assistance (Li et al. , 2024a ) , customized feedback (Fu et al. , 2024 ) , and adaptive learning paths (Lee et al. , 2024 ) . Such features empower learners to regulate their cognitive processes and support various tasks, including self-reflection (Li et al. , 2023 ) , goal-setting (Deng et al. , 2023 ) , and problem-solving (Kumar et al. , 2023 ) . This makes LLM-based educational agents particularly valuable in self-learning contexts, where individualized and adaptive support is critical. Self-directed learning is a process involving continuous cognitive and behavioral adjustment aimed at understanding specific knowledge domains (Demirbag, 2021 ; Huang et al. , 2023 ) . Learning outcomes can vary depending on individual strategies, and pedagogical research emphasizes the importance of active engagement and self-regulation in this process (Wang et al. , 2015 ; Wekerle et al. , 2022 ) . Recent HCI studies have explored how LLM-based educational agents can enhance the effectiveness of self-learning by promoting autonomy, particularly from students’ perspectives (Limna et al. , 2023 ; Li et al. , 2024b ) . For instance, Goslen et al. ( Goslen et al. , 2024 ) proposed an LLM-based system that automatically generates student planning strategies based on historical interaction data, enabling dynamic support in game-based learning environments. Through fine-tuning for domain-specific instruction, these agents have demonstrated success in areas such as language learning (Liu et al. , 2024c ) , technology literacy (Gao et al. , 2024 ; Hartley et al. , 2024 ) , and research ideation (Liu et al. , 2024b ) . These advances enable self-learners to effectively manage their cognitive processes and improve practical skills with reduced dependence on human guidance. Despite these advances, LLM-based educational agents have yet to achieve the dynamic reciprocal interaction where learners and systems mutually shape each other’s development, as envisioned by pedagogical scholars (Schulte and Budde, 2018 ) . This limitation stems primarily from the agents’ inability to guide users in crafting effective prompts (Shah and Bender, 2022 ; Fiannaca et al. , 2023 ; Zamfirescu-Pereira et al. , 2023 ) , which increases cognitive load and hinders knowledge acquisition. Although features such as automatic question generation (Zheng et al. , 2024b ; Hu et al. , 2024 ) and dialogue history management (Safranek et al. , 2023 ) improve interaction intuitiveness, a lack of systematic guidance limits the ability of self-learners—especially those unfamiliar with AI or interactive learning—to fully harness these agents’ potential. Our study addresses this gap by investigating how different mental models shape self-learners interaction patterns with LLM-based educational agents. These insights will inform the design of more effective, accessible, and inclusive AI-enhanced educational systems, ultimately enhancing learner engagement for a broader audience. 2.2. Integrating Epistemology for LLM Sensemaking The advent of LLMs has spurred their integration into tools augmenting human sensemaking during self-learning. Sensemaking is a vital cognitive process that influences how learners process information, ultimately impacting their learning outcomes and efficiency ( Marchionini , 2019 ; Rong et al. , 2023 ) . This cyclical process encompasses several cognitive stages: acquiring , comprehending , integrating , and applying new knowledge (Pirolli and Card, 2005 ) . Recognizing LLMs’ potential to transform this process, the HCI research community has explored various approaches to enhance the different cognitive stages (Lee and Ma, 2024 ; Gero et al. , 2024 ; Ma et al. , 2024 ) . For instance, Zheng et al. (Zheng et al. , 2024b ) enhanced knowledge search and acquisition through an interdisciplinary literature search system that facilitates cross-domain synthesis. Building on this foundation, Suh et al. (Suh et al. , 2023 ) advanced knowledge comprehension by developing an interactive system that reveals multilevel abstractions of complex information spaces. Taking a different approach, Jin et al. (Jin et al. , 2024 ) focused on knowledge construction by leveraging LLMs as teachable agents within a “learning by teaching” paradigm. Despite their promise, existing LLM sensemaking applications rarely consider how individual epistemological schemas shape knowledge construction. Epistemology–the philosophical study of knowledge and belief that guides learners to evaluate the credibility and relevance of information sources–underpins how learners make sense of new information ( Hofer and Pintrich , 1997 ; Schommer-Aikins , 2004 ; Shirzad et al. , 2022 ) . Pedagogical research emphasizes that the effectiveness of sensemaking depends fundamentally on learners’ epistemological schemas–their structured beliefs about knowledge’s nature , source , and justification ( Hofer and Pintrich , 1997 ; Bråten , 2010 ; Baker and Anderman , 2020 ) . These beliefs critically shape how individuals process information during learning, from initial interpretation to final evaluation ( Zhang and Soergel , 2020 ) . For example, Baker et al. ( Baker and Anderman , 2020 ) demonstrate that such personal epistemological schemas influence the entire spectrum of self-learning, affecting learners’ motivation, achievement levels, and engagement with learning materials. Given the fundamental role of epistemology in knowledge construction, prior research has investigated applying epistemological frameworks to AI models for educational tools ( Woolf et al. , 2013 ; Zhang et al. , 2023 ) . For instance, Rosé et al. ( Rosé et al. , 2008 ) proposed an AI-driven framework grounded in an understanding of epistemology, which can trace learners’ knowledge construction processes, thereby improving learning outcomes. However, the rapid adoption of LLMs has shifted recent scholarly attention toward their capacity to disrupt traditional teaching practices and introduce epistemic paradoxes ( Sibilin , 2023 ; Chavanayarn , 2023 ; Cassinadri , 2024 ) . Consequently, the potential of leveraging LLMs for accommodating learners’ epistemological schemas, which is crucial for sensemaking, remains largely unexplored. Our research addresses this gap by integrating epistemological frameworks into LLM-based systems through prompt engineering, guiding users through the processes of problem framing, evidence evaluation, and reflection, key stages of sensemaking. This domain-agnostic approach ensures versatility and offers a structured yet adaptable framework for knowledge exploration. 2.3. Toward Effective Learner-LLM Interactions Understanding and improving the interactivity in learning activities remains a persistent focus in pedagogy. The Interactive , Constructive , Active , and Passive (ICAP) framework, proposed by Chi (Chi, 2009 ) , categorizes learner engagement modes based on observable behaviors and underlying cognitive processes, providing a foundation for understanding learner interactivity. A subsequent five-year study (Chi and Wylie, 2014 ) demonstrated the relative effectiveness of these engagement modes, ranking them from most to least effective: interactive, constructive, active, and passive. This finding has fueled further learner-centered research exploring the relationship between engagement modes and learning outcomes in specific educational scenarios (Wang et al. , 2015 ; Schulte and Budde, 2018 ) . While some variations exist in the resulting classifications and analyses of contributing factors, the principle that “ good learning is active learning ” remains widely accepted (Brown et al. , 2014 ) . However, translating the ICAP framework into practice, particularly in designing interactive and constructive activities, has proven challenging (Chi et al. , 2018 ) . Advances in technology create new opportunities to apply the ICAP framework. Recent educational research has extended the ICAP theory to the context of TEL (Sailer et al. , 2021 ; Wekerle et al. , 2022 ; Sailer et al. , 2024 ) . Building on the ICAP framework, Sailer et al. (Sailer et al. , 2024 ) introduced the Substitution , Augmentation , Modification , and Redefinition (SAMR) model to classify levels of technology integration in learning. The introduction of LLM-based educational agents marks a transformative step toward SAMR’s “redefinition” level, especially in self-learning scenarios. These agents enable scalable conversational interactions that can enhance learner engagement. Although the conversational interface of LLM-based educational agents suggests simplicity and intuitive interaction similar to human communication (Smestad and Volden, 2019 ) , extant HCI literature reveals notable disparities in how learners benefit from these tools (Chen et al. , 2024a ) . This variability often arises from learners’ diverse motivations, levels of LLM literacy, mental models, and educational backgrounds, ultimately influencing their interaction strategies and learning outcomes (Mogavi et al. , 2024 ; Zamfirescu-Pereira et al. , 2023 ) . For instance, research on help-seeking behaviors among self-learners highlights how self-efficacy, perceived task difficulty, and tool accessibility influence their resource utilization (Moores and Chang, 2009 ; Karabenick and Dembo, 2011 ; Glassman et al. , 2015 ; Urgo and Arguello, 2022 ) . Overconfident learners, in particular, may only seek external support when facing substantial challenges or when perceiving the tool as beneficial for their understanding (Moores and Chang, 2009 ) . This tendency can hinder their ability to accurately assess their learning progress. Furthermore, novice AI users often struggle to craft effective prompts during LLM interactions, compounding the challenges they face (Zamfirescu-Pereira et al. , 2023 ; Fiannaca et al. , 2023 ) . To promote effective learner-LLM interaction, designing more inclusive LLM-based tools that provide effective cognitive support and accommodate the diverse needs of self-learners is essential (Cohn et al. , 2024 ) . Drawing on the ICAP framework (Chi and Wylie, 2014 ) and our previously discussed epistemological perspective, our research examines how self-learners interact with LLM-based educational agents in practice. By analyzing these interaction patterns and their underlying mental models, we explore how to enhance LLM prompt engineering to better facilitate learner interactivity and knowledge co-construction. 3. Formative Study This section describes our formative study, which adopts a user-centered approach (Vredenburg et al. , 2002 ) . The study investigates participants’ interaction behaviors , epistemological schemas , and the challenges they face during self-directed learning. By synthesizing these dimensions, we derive a categorization of participants’ interaction patterns and deduce the latent mental models that guide their engagement when learning new topics. These findings inform the design priorities of our system. Specifically, we outline the recruitment process and participant demographics, describe the study procedure, and explain the data analysis methods. This study was conducted in accordance with all relevant ethical guidelines and received full approval from our university’s Institutional Review Board (IRB). Prior to data collection, verbal informed consent was obtained from all participants. 3.1. Recruitment and Participants Recruitment We employed a dual-pronged approach, combining convenience and snowball sampling strategies, to gather a representative sample of target users who use LLM-based chatbots for self-learning. We disseminated recruitment messages with contact information through multiple channels, including social media, word of mouth, and campus bulletin boards at five institutions of higher education in East Asia and North America. Our recruitment criteria consisted of the following: (1) adults aged 18 years or older, (2) individuals capable of interacting with LLM-based chatbots, (3) those with at least three months of prior experience using LLMs for self-learning, and (4) participants without prior knowledge of blockchain technology, which served as the primary reading material for our self-learning task. We received a total of 35 applications and conducted a thorough screening of respondents based on our criteria. We stopped the recruitment process when data analysis results reached theoretical saturation (Clarke and Braun, 2017 ) , meaning no new insights emerged from participants’ dialogue records or semi-structured interviews (SSI). Ultimately, we recruited 26 participants with varied educational backgrounds, from various disciplines, and from seven countries and regions (see Table 6 in Appendix A ). Table 1. Descriptive statistics of participants in the formative study. This diverse sampling pool provides insights into design requirements for improving LLM-based chatbots for self-learning across various demographics. Participants The 26 participants were between 23 and 32 years of age (M = 27.04, SD = 2.28), and carefully selected to represent diverse backgrounds and experiences with LLM-based tools (see Table 1 ). Sixteen held a master’s degree, followed by four with bachelor’s degrees and six with doctorates, indicating a strong representation of individuals engaged in advanced studies or research. Our participants represented a diverse range of disciplines, with a substantial portion (13) coming from STEM fields–a group shown to have positive perceptions of LLMs as learning tools (Bernabei et al. , 2023 ) . The remaining participants were from social sciences (4), interdisciplinary studies (5), and other fields (4), ensuring broader perspectives. Most participants (22) had experience in academic training, providing a foundation for evaluating LLM-based self-learning tools. While ten participants had 1-2 years of experience using LLM tools, a significant portion (15) had been using them for less than a year, and only one for over two years, indicating a mix of early adopters and more recent users. Finally, participants’ familiarity with LLM-based chatbots varied: 21 reported being somewhat (11), very (6), or extremely (4) familiar, and five were unfamiliar. This distribution enabled comparison between more and less adept chatbot users, a crucial aspect of our study design. 3.2. Formative Study Design The formative study aimed to investigate participants’ interaction behaviors , epistemological schemas , and the challenges they encountered during self-learning through a task-focused probe system. Before formally launching the formative study, a pilot study was conducted with two team members from diverse disciplinary backgrounds to validate the process and estimate the required duration. 3.2.1. Study Setup The formative study allowed both in-person (n=12) and online (n=14) participation. In-person sessions were held at the recruiting institutions. Online sessions were conducted via Zoom 2 2 2 https://zoom.us/ to facilitate real-time interaction and observation of participant behavior. To provide consistent participant experiences within a controlled learning environment, we employed Qualtrics 3 3 3 https://www.qualtrics.com/ for online surveys and a task-focused probe system accessed via the Web directly . This system included raw learning materials, a textual article, and an LLM-chatbot limited to responding only to the provided learning material (see Fig. 10 in Appendix A.2 ). This controlled setup ensured a fair self-learning process, free from misinformation and external online sources. Aiming for smooth operation, participants received essential documents, including the consent form and study guidance, one day prior to their scheduled session. These documents detailed the research goals and experimental procedures and addressed any potential questions. Notably, we did not share the probe system link in advance to avoid pre-exposure to the learning materials, which could potentially bias our assessment of user learning performance. The formative study sessions were allocated 90 minutes based on the results of the pilot study. Figure 2. Formative Study Workflow (N=26): A three-phase investigation of LLM interaction patterns during self-learning. (A) Data Collection: Gathered dialogue records, quiz scores, survey responses, and transcripts of semi-structured interviews. (B) Thematic Analysis: Classified four distinct LLM interaction patterns (proactive, validation-seeking, content-focused, receptive). (C) Epistemological Analysis and Design Implications: Compared and integrated epistemological frameworks across the identified interaction patterns, revealing insights into participant challenges and informing the design requirements for an epistemologically-grounded LLM-assisted learning system. 3.2.2. Study Procedure Upon receiving informed consent, the study proceeded with the following steps (Fig. 2 A): Self-Learning Task Participants initially completed a sensemaking task using an informational article on non-fungible tokens (NFTs) (Creighton, 2023 ) . Using a think-aloud protocol, participants provided real-time feedback on the probe system’s design and usability during their interactions. Participants self-regulated their learning pace, ceasing interaction once they felt they had adequately understood the article. They then downloaded a record of their dialogue with the LLM for future reference. Task completion time ranged from 15 to 35 minutes (M = 20.21, SD = 5.12). Quiz Participants then completed a 15-minute, ten-point quiz assessing the comprehensiveness and accuracy of their acquired knowledge. The quiz, consisting of four multiple-choice and two short-answer questions, was designed by two NFT domain experts who also independently graded the responses. The final scores were the average of the two experts’ grades. Survey and Semi-Structured Interview To further understand individual learning experiences, a seven-point Likert scale survey was administered, gathering data on LLM usage patterns (e.g., familiarity, modalities) and perceived challenges. Semi-structured interviews then explored individual learning processes in depth, with participants sharing the rationale behind their query chains, challenges, and suggestions for improving LLM interactions. 4 4 4 For illustrative examples of conversation records corresponding to each interaction pattern, please refer to the supplementary materials. Each interview lasted between thirty minutes and one hour, with audio recordings and verbatim transcriptions conducted. Figure 3. This figure presents our six-step process for analyzing participants’ interaction behaviors and epistemological frameworks based on data collected from our formative study. 3.3. Data Analysis We employed reflexive thematic analysis (Braun and Clarke, 2019 , 2021 ) to explore participants’ sensemaking processes and categorize their interaction patterns. Building on existing literature (Aliannejadi et al. , 2021 ; Schneider et al. , 2023 ) , we defined interaction patterns based on two key aspects. The first focuses on participants’ interaction behaviors , as reflected in the queries they generated during the self-learning process. The second dimension pertains to their epistemological schemas , derived from their chain of queries and self-reported knowledge construction processes detailed in the interview transcripts. We also analyzed the challenges participants encountered while interacting with the LLM through a detailed examination of the interview transcripts, which was further complemented by their survey responses. The complete process and use of collected data sources are illustrated in Fig. 2 B. Given the exploratory nature of our study within the emerging field of LLM-mediated self-learning, we adopted a predominantly inductive approach. This approach enabled us to develop a query taxonomy of learner-LLM interactions while drawing on an epistemological framework (Aksan, 2009 ; Mokhtari, 2014 ; Huang et al. , 2023 ) to analyze participants’ knowledge construction processes. This flexible approach enabled us to explore the multifaceted nature of user experiences in depth (Guest et al. , 2011 ) , while grounding our analysis in relevant theoretical constructs. Three analysts participated: two authors (who observed all self-learning tasks and conducted all interviews) and one NFT domain expert. 3.3.1. Interaction Behaviors Analysis Initially, the three analysts independently examined participants’ interaction behaviors within the self-learning task (see Fig. 3 A). The goal was to systematically categorize 179 participant-generated queries based on the knowledge dimensions explored and the conversation strategies employed. The analysts met regularly to share coding results, discuss discrepancies, and refine the codebook until reaching a consensus for the query taxonomy. The resulting query taxonomy not only highlights the types of queries generated by participants but also captures the frequency of their interactions and the diversity of conversation strategies they employed. Based on this analysis, participants’ interaction behaviors were mapped along a spectrum ranging from prompt-naive (low interaction frequency; few exploration dimensions and conversation strategies) to AI-interactive (high interaction frequency; diverse exploration dimensions and conversation strategies) (see Fig. 3 B). To further contextualize these patterns, the analysts conducted a detailed examination of each participant’s chain of queries (see Fig. 3 C). This examination informed the subsequent semi-structured interviews designed to elicit a nuanced understanding of participants’ epistemological schemas for self-learning. 3.3.2. Epistemological Schemas Analysis To investigate participants’ epistemological schemas, we integrated the previously established query taxonomy, the analysis of query chains, and the data gathered from the semi-structured interviews. During these interviews, we asked participants about their information needs and how they constructed meaning, referencing their specific query flows from the dialogue records. We analyzed the interview data (22.5 hours recorded, transcribed into 26 transcripts) alongside the corresponding dialogue records. This analysis employed a three-dimensional framework focusing on the nature, source, and justification of knowledge to understand participants’ personal epistemological schemas (Aksan, 2009 ; Mokhtari, 2014 ; Huang et al. , 2023 ) . Guided by this framework, we systematically coded the interview data within Atlas.ti 5 5 5 https://atlasti.com/ software (see Fig. 3 D). We met weekly to review findings, resolve disagreements, and refine the codes until a consensus was reached. As a result, we placed participants’ epistemological schemas along a spectrum from reflective-minded to confirmatory-oriented (see Fig. 3 E). Reflective-minded patterns are characterized by a dynamic view of knowledge, knowledge construction through exploration, and justifications through multi-faceted validation. In contrast, confirmatory-oriented patterns demonstrate an authoritative view of knowledge, reliance on external expertise, and justification by seeking standard answers. 3.3.3. Interaction Patterns Classification This mixed-methods approach revealed two key spectra for categorizing participants based on a two-dimensional analytical framework: interaction behavior , ranging from prompt-naive to AI-interactive, and epistemological schema , spanning from reflective-minded to confirmatory-oriented. The integration of these dimensions provided a robust and holistic analysis, generating a quadratic classification of interaction patterns (similar to (He et al. , 2023 ) ): proactive , validation-seeking , content-focused , and receptive (see Fig. 3 F). Figure 4. This 2x2 matrix presents four distinct interaction patterns, categorized by AI interactivity and reflective-mindedness: Prompt-Naive, Confirmatory-Oriented, Reflective-Minded, and AI-Interactive. This classification reflects the mental models associated with different learning strategies observed during self-learning. 4. Findings This section unpacks four interaction patterns and their epistemological differences. These analyses, combined with survey responses and interview data, revealed key challenges and informed the design requirements for our LLM-supported self-learning prototype (see Fig. 2 C). 4.1. Learner-LLM Interaction Patterns This section illustrates four learner-LLM interaction patterns. These patterns are characterized by their query types, conversational strategies, self-reported learning workflows (gathered through semi-structured interviews), and quiz performance. 4.1.1. Proactive (n = 8) Eight participants exhibited the proactive interaction pattern marked by high interactivity and reflectiveness and actively pursued knowledge (see Fig. 4 A). They viewed the chatbot as a collaborative partner, engaged in extended, multi-round dialogues, and critically examined responses. The proactive interaction pattern is demonstrated through the participants’ elevated query volume and diverse exploratory query types (see Table 2 ). They delved deeper into topics by asking about “attributes of concepts,” “co-existent concepts,” “realistic application,” and “real-world consequences”–all highlighting a desire to understand a topic from multiple angles. For example, P9 explained, “ I don’t just want the ‘what,’ I want the ‘why’ and the ‘how does this connect to other things?’ ” In addition, these participants strategically employed queries such as “change perspectives” and “rephrase requests,” actively guiding the conversation toward providing a more holistic understanding. For instance, P12 revealed, “ I rephrase and ask about the same concept multiple times in different ways to get a more comprehensive and specific response, instead of an overly general one. ” This demonstrates an active role in shaping the learning experience. This interaction pattern was often reflected in their high quiz scores, ranging from 7 to 9 out of 10 (M = 8.28, SD = 0.62), and suggested active learning strategies driven by curiosity and a desire for deeper comprehension. Table 2. This table presents the query taxonomy developed during the formative study, structured around three key aspects: Knowledge Dimension , Conversation Strategy , and Interaction Frequency (visualized numerically via a heat map). Questions marked with an asterisk (*) represent follow-up queries designed to provide deeper insights. While the heat map directly quantifies the AI-interactivity dimension, reflective-mindedness—captured through triangulated interview analysis—is only partially represented here. 4.1.2. Validation-Seeking (n = 6) Six participants displayed the validation-seeking interaction pattern, characterized by high interactivity but low reflectiveness, primarily engaging with chatbots to confirm existing knowledge (see Fig. 4 B). Their high-frequency interaction behaviors primarily involve asking specific questions with predetermined expectations to validate assumptions or test hypotheses (see “validate hypothesis” in Table 2 ) Their repeated queries about “co-existent concepts” and “realistic applications” reflected this desire to align new information with established understanding and real-world scenarios (see Table 2 ). As P1 noted, “ Relating new concepts to hypotheses formed from my existing knowledge boosts my confidence in self-learning. ” Furthermore, they often apply conversational strategies such as “validate hypothesis” and “assess accuracy,” emphasizing their preference for reinforcing existing knowledge rather than exploring new perspectives. As P22 shared, “ Instead of constantly trying to synthesize vast new concepts, presenting a hypothesis and seeking clear feedback is more straightforward. ” Their quiz scores, ranging from 5 to 7.5 out of 10 (M = 6.33, SD = 0.98), suggest that while these participants prioritize accuracy, relying on their existing understanding may not always facilitate the necessary expansion of their knowledge base for a thorough grasp of the subject. 4.1.3. Content-Focused (n = 6) Six participants demonstrated the content-focused interaction pattern, characterized by low interactivity but high reflectiveness in their behaviors, prioritizing thorough understanding over fragmented knowledge (see Fig. 4 D). These participants often engaged in extensive exploration of learning materials before directly engaging with the LLM chatbot. As P14 articulated, “ I do not want to ask superficial questions. I need time to process the information and figure out valuable questions .” With this pattern, participants tended to use the LLM less often, primarily for retrieving information directly related to specific aspects of the learning materials, such as “components of concepts,” “co-existent concepts,” and “realistic applications.” This pattern is further evidenced by their sole use of the “summarize content” strategy (see Table 2 ). P17 aptly captured this tension, noting, “ Simultaneously grasping new concepts and forming insightful questions is challenging, so I mainly use the LLM to facilitate understanding .” However, this approach may not always translate to high performance within a time-limited assessment context, as suggested by the significant individual variability in quiz scores, ranging from 5 to 8 out of 10 (M = 6.58, SD = 1.64). 4.1.4. Receptive (n = 6) The receptive interaction pattern was evident in six participants, whose behaviors were marked by minimal engagement, limited reflectiveness, and passive interaction with the LLM chatbot (see Fig. 4 C). Accustomed to traditional learning methods like reading, online courses, and structured video lectures–which typically involve limited interaction–these participants often struggled to formulate precise questions or to fully leverage the chatbot’s capabilities. As P13 admitted, “ I am not used to learning with an LLM. I do not think asking questions is necessary. ” This sentiment is further evidenced by a lack of discernible patterns in their query topic and strategy preferences, indicating a less systematic LLM-interaction approach (see Table 2 ). Furthermore, their quiz scores, ranging from 5 to 7 out of 10 (M = 5.75, SD = 0.82), generally suggest that those who struggle to adapt to this interactive learning modality may experience less favorable learning outcomes. 4.2. User-Centered Design This section illuminates the epistemological differences across the four learner-LLM interaction patterns. Integrating insights from the surveys and interviews, this analysis reveals the primary challenges ( C1-3 ) participants face when using LLMs and informs design requirements ( DR1-5 ) for enhancing LLM-supported self-learning tools. Table 3. This table presents the matrix analysis of epistemological differences across four interaction patterns based on the three-dimensional epistemological framework. 4.2.1. Epistemological Differences in LLM Interactions A comparative analysis of the four interaction patterns, presented in Table 3 using a matrix method (Groenland, 2018 ) , reveals notable epistemological differences. This analysis, informed by the previously established three-dimensional epistemological framework (Hofer and Pintrich, 1997 ; Aksan, 2009 ; Mokhtari, 2014 ; Huang et al. , 2023 ) described in § 3.3.2 , considers the perceived nature of knowledge (absolute vs. evolving), the primary source of knowledge (external authorities vs. active construction), and the criteria for justifying knowledge. The contrast between proactive and receptive interaction patterns revealed fundamentally different mental models of sensemaking. Proactive patterns, marked by actively seeking diverse perspectives and critical reflection, facilitate dynamic knowledge construction and validation through iterative interaction and exploration. While this iterative exploration promotes deep understanding, it creates a practical challenge of efficiently verifying the extensive information generated through LLM interactions (see C1 in § 4.2.2 ). In contrast, receptive patterns rely heavily on external authority and a more passive way of engaging, which limits effective knowledge construction with LLMs. As a result, fostering meaningful interactivity and learner agency (see C2 in § 4.2.2 ) becomes particularly challenging when users treat LLMs primarily as information sources rather than interactive learning partners. Participants exhibiting validation-seeking and content-focused interaction patterns occupy an interesting middle ground. While they value accuracy and structure, their strategic use of LLMs for targeted information suggests an openness to expanding their understanding beyond curated materials. However, these patterns tend to emphasize alignment with authoritative sources and validation of prior knowledge. They may inadvertently limit learners’ exploration of diverse perspectives and create a gap between their desire for comprehension and reliance on established sources (see C3 in § 4.2.2 ). 4.2.2. Challenges in LLM Interactions Our analysis of epistemological differences across different LLM interaction patterns, integrating survey (see Fig. 5 ) and semi-structured interview results, reveals unique challenges participants encountered during LLM-mediated self-learning. C1 . Inefficient Information Verification. Across various interaction patterns , effectively and efficiently verifying the accuracy of LLM-generated information poses a significant challenge (Cucuiat and Waite, 2024 ) . Participants with proactive and validation-seeking interaction patterns find the validation process time-consuming and cumbersome, hindering their ability to assess the reliability of AI’s responses. For instance, P9 commented, “ ChatGPT sometimes fabricates information, so I have to verify its explanations against multiple sources, which can takes longer than finding the answer myself. ” C2 . Limited Interactivity and Agency. The interactive nature of LLMs presents challenges to self-learners accustomed to more structured or passive approaches (Zamfirescu-Pereira et al. , 2023 ) . Participants showing content-focused interaction patterns, find that the lack of seamless integration between LLM responses and their preferred learning materials diminishes their sense of control over the learning process. As P11 noted, “ Switching between the chat and readings is frustrating. Especially when the AI’s responses do not align with my learning goals–it just throws me off track. ” Similarly, participants with receptive tendencies struggle to adapt to the open-ended nature of interacting with LLMs, finding it difficult to formulate effective queries and manage the volume of information received. As P24 explained, “ Asking good questions is challenging in itself–you need to know the topic already. Helpful prompts from the AI would be great; otherwise, I am lost. ” C3 . Confirmation Bias. Participants showing validation-seeking preferences are particularly vulnerable to confirmation bias. They use LLMs to reinforce pre-existing beliefs rather than critically examining alternative perspectives or challenging their own assumptions. For example, P22 admitted, “ I sometimes keep tweaking my prompts until ChatGPT gives me what I expected, even if that means ignoring other perspectives. ” This behavior underscores how unstructured LLM interactions can narrow learning pathways. It highlights the need for systematic query approaches and metacognitive support to help learners examine their thinking processes and develop more comprehensive exploration strategies. Figure 5. A substantial portion of participants reported challenges using the LLM-supported chatbot for self-study, particularly with question formulation, response clarity, and the achievement of deep understanding. 4.2.3. Design Requirements The challenges uncovered in our formative study directly informed the five key design requirements for our LLM-supported self-learning system: DR1 . Validated Information and Validation Support ( C1 ). Systems must provide readily validated information or offer user-friendly methods for verifying the accuracy of LLM outputs. This includes facilitating precise cross-referencing between LLM responses and the original source material. DR2 . Smooth Onboarding and Reduced Cognitive Load ( C2 ). Learners expressed a preference for LLMs to autonomously generate systematic follow-up questions, minimizing the initial interaction effort. This approach reduces the barrier to entry and enhances self-study efficiency by reducing the time spent crafting prompts. DR3 . Sustained Engagement and Deepened Exploration ( C2 ). Engaging and fluid dialogue is crucial to maintaining users’ engagement while guiding them toward a deeper understanding of new concepts. Supporting flexible, multi-round dialogues enables learners to progressively deepen their understanding of concepts through both horizontal and vertical exploration. DR4 . Fostering Openness and Multidimensional Thinking ( C3 ). Comprehensive knowledge acquisition requires learners to consider multiple perspectives and generate diverse hypotheses. A well-designed taxonomy of queries, informed by a robust epistemological framework, can effectively support this type of multifaceted thinking. DR5 . Facilitating Metacognition and Reflection ( C3 ). Systems should support metacognition by structuring the interaction process and encouraging reflection on learning progress and comprehension levels. Recording and visualizing the user’s chain-of-query can enhance knowledge retrieval efficiency and help identify potential gaps in understanding. 5. CausaDisco We present CausaDisco , an epistemologically-informed system designed to enhance sensemaking during self-learning. Carefully designed features enable CausaDisco to encourage user interaction with an LLM chatbot to enhance sensemaking, fulfilling the design requirements outlined in § 4.2 . In the following sections, we illustrate CausaDisco ’s interface design, highlighting its integration of the epistemological framework and prompt engineering techniques. 5.1. Interface Design CausaDisco integrates four coordinated views to streamline the self-learning process within complex educational content. The Embedded Content View (Fig. 6 A) provides users with access to the original learning materials, allowing immediate verification and reference ( DR1 ). Complementing this, the Concept Graph View (Fig. 6 B) visually maps out the core concepts and their interrelationships, clearly and concisely representing the structure of the material ( DR2 ). The interactive Q&A Conversation View (Fig. 6 C) enables users to interact with the chatbot and to choose from a range of suggested follow-up questions, fostering a deeper exploration of specific topics ( DR3, DR4 ). Lastly, the Tree Map View (Fig. 6 D) organizes the user’s inquiry process into a logical framework, facilitating easy retrospection and effectively tracking their learning progression ( DR5 ). Figure 6. System Overview: The interface features four main views: A) Embedded Content View ; B) Concept Graph View ; C) Q&A Conversation View ; D) Tree Map View . The backend comprises two main components: (E1) The Four Causes principles, which inform the prompt engineering for generating follow-up questions (FQ); and (E2) A logging function that records the user’s chain of queries (E3) during interactions with the LLM chatbot. 5.1.1. Embedded Content View The Embedded Content View (Fig. 6 A) has two sections. The right-hand side displays the original learning materials, including hierarchical headings, author information, publication date, main text, and illustrations (when available) (Fig. 6 A2). The left-hand side provides an interactive navigation bar for users to locate sections of interest swiftly and can be hidden to maximize screen space (Fig. 6 A1). This design facilitates users in verifying the authenticity or accuracy of LLM responses ( DR1 ). They can effortlessly use the navigation bar to pinpoint the pertinent section in the original text to cross-reference information and delve into more details. Justifications: The design of the Embedded Content View prioritizes text-based learning resources, as informed by our formative study findings. This focus on text enables seamless AI integration without disrupting learning flow while allowing efficient content parsing for accurate, context-aware AI responses. Text-based content also offers flexible presentation and interaction, accommodating diverse interaction patterns . 5.1.2. Concept Graph View The Concept Graph View (Fig. 6 B) comprehensively summarizes the core concepts and their interrelationships within the learning materials. We employ a semi-automated approach to enhance the accuracy and granularity of the concept graph. By first identifying high-frequency keywords in the materials, we established the scope of the concept graph by incorporating a glossary of concepts provided by ChatGPT-4. After analyzing the original texts with assistance from two domain experts in the respective fields (in our case, NFTs and semiotics), we identified four types of relationships between the key concepts: foundational prerequisites, defining traits, illustrative examples, and the influence of one concept’s functions or characteristics on another (Fig. 6 B2). Accordingly, we created the corresponding concept graphs for different learning materials (Fig. 6 B1). This view offers users a concise overview of the knowledge landscape, using keywords as search indices to create a straightforward starting point for interacting with the LLM ( DR2 ). 5.1.3. Q&A Conversation View The Q&A Conversation View (Fig. 6 C) equips users with epistemologically-informed follow-up questions during their interactions with the LLM chatbot. This view autonomously generates four alternative follow-up questions immediately after the chatbot responds to the previous query (Fig. 6 C2). These follow-up questions support the deep exploration and comprehension of the concepts discussed, enriching the experience beyond the typical single-turn conversation ( DR3 ). Users can interact with any of the automatically generated follow-up questions by clicking on them to obtain a response from the LLM chatbot. They can also customize a question by clicking the “modify” button next to it. This button allows them to edit the question in the input box before submitting it. The follow-up questions are crafted to guide users in exploring the topic from multiple perspectives ( DR4 ), fostering a more holistic understanding. The underlying mechanism for generating these questions is discussed in detail in § 5.2 . This view also logs all dialogue interactions between users and the LLM chatbot (Fig. 6 E2). 5.1.4. Tree Map View The Tree Map View (Fig. 6 D) dynamically visualizes users’ query histories, allowing for a quick review of their learning progress and facilitating timely adjustments to their learning strategies ( DR5 ). This view is characterized by its cross-view interactions. When users interact in the Q&A Conversational View , the Tree Map View simultaneously and automatically records the “parent” questions they ask. In a hierarchical node-link format, it presents the “child” follow-up questions selected by users (Fig. 6 E3). The query tree guides users to progressively explore the topic in depth until a new question is posed, indicating the end of the learning session for the previous parent topic. At this point, the view automatically creates a new tree map for the new parent topic. Users can freely expand or collapse branches associated with the same parent topic to adjust the level of detail shown in the tree map. They can also adjust the tree map’s size and position through scrolling and panning. This view preserves all generated tree maps, assisting users in managing their interaction records and reviewing their logical structures, supporting reflection on the learning process, and adjusting their question strategies. Justifications: We also considered using a table format to display users’ query histories. However, feedback from participants during our formative study suggested that graphs are more intuitive than tables, as they better illustrate the hierarchical relationships between questions. 5.2. Prompt Engineering This section details the process of generating the follow-up questions outlined in § 5.1.3 . We discuss the epistemological framework we selected, the rationale behind its choice, and how it is incorporated into our prototype. 5.2.1. Leveraging the Epistemological Framework As discussed in § 4.2 , we aim to design follow-up questions that facilitate both comprehensive exploration and multidimensional thinking. Assigning the LLM, ChatGPT in our case, the role of a tutor seems appropriate; however, this poses the challenge that the generated questions might merely be relevant without addressing the user’s understanding gaps. Furthermore, ChatGPT’s tendency to mimic expert vocabulary (Zollman et al. , 2023 ) may be unsuitable for our application scenario. Therefore, we equip ChatGPT with a structured framework to guide its output, ensuring the follow-up questions provide effective cognitive support for self-learners. Given that the follow-up questions are designed to facilitate a comprehensive exploration of the topic, incorporating this framework into the prompt should yield questions that span all categories we summarize in Table 2 . Furthermore, our approach needs to be domain-agnostic and easily integrated into prompts. Consequently, we opted for an established theory that ChatGPT understands and incorporated it as a guiding term in our prompts. After exploring various options, we found that popular learning theories from behaviorism or cognitivism were inadequate for our needs. Additionally, studies on student mental models often focus on specific domains (Fratiwi et al. , 2020 ; Taylor et al. , 2003 ) . To address this, we turned to epistemology, which centers on acquiring knowledge and understanding reality. Epistemological theories have long been integrated into pedagogical practices to foster critical thinking and active engagement in exploring the unknown (Duschl et al. , 1990 ; Macallister, 2012 ; Kotzee, 2018 ) . Among various epistemological theories, Aristotle’s Four Causes framework ( Material , Formal , Efficient , and Final Causes) (Hocutt, 1974 ) stands out for its ability to refine cognitive processes (Pérez and Ziemke, 2007 ) and improve educational outcomes (Taylor and Sondermeyer, 2023 ) . Specifically, this framework offers a comprehensive analytical lens for understanding the essence of any entity, covering all question types summarized in Table 2 . We used a deductive method to map the connections between our query taxonomy and the Four Causes framework (Table 4 ). The material cause naturally corresponds with the “components of concepts” category, which refers to the fundamental elements constituting a concept. The formal cause, which describes the distinct form or arrangement defining an object, aligns with the “attributes of concepts” category. Additionally, the formal cause includes “co-existent concepts,” helping clarify the characteristics that distinguish a concept from others. The efficient cause, representing the driving force behind an entity’s existence or change, aligns with both “realistic application” and “development of concepts,” as these categories explore the motivations and mechanisms of change. Lastly, the final cause, which denotes the purpose or intended outcome of an entity, maps to the “significance of concepts” and “real-world consequences” categories (Falcon, 2006 ; Cohen and Reeve, 2000 ; Charlton and others, 1983 ) . Table 4. The mapping between the epistemological framework, i.e., Four Causes , and the query taxonomy identified from the dialogue data generated by participants during the self-learning task. Four Causes Original Definition Query Types Examples Material Cause “ That out of which a thing comes to be and which persists. ” Components of Concepts “Why must NFTs be involved with blockchain?” “Can users mint NFTs without any assets?” Formal Cause “ The form or the archetype, i.e. the statement of the essence. ” Attributes of Concepts “What is the meaning of non-fungible?” “Do NFTs protect copyrights?” Co-existent Concepts “What are the benefits of NFTs compared to cryptocurrencies?” “What are the relationships between NFTs, ETH, and blockchains?” Efficient Cause “ The primary source of change and the various factors that contribute to that change. ” Realistic Application “How are Non-fungible tokens (NFTs) regulated?” “What are the uses of NFTs beyond investigation and collection?” Development of Concepts “How was the first NFT created?” “What is the outlook for NFTs?” Final Cause “ In the sense of end or ‘that for the sake of which’ a thing is done. ” Significance of Concepts “What is the importance of NFTs?” “Why do we need NFT as a newly-emerging technology?” Real-world Consequences “What are the benefits of owning an NFT?” “What is the significance of NFTs in the current capitalist society?” 5.2.2. Integrating Epistemological Framework Here we elucidate the process of generating follow-up questions using the Four Causes framework within our system, leveraging the capabilities of ChatGPT. It is worth noting that while a foundational understanding of the Four Causes is not a prerequisite for users, the questions are strategically designed to guide their inquiry towards more complex and contextually rich content. In our initial evaluation, we assessed ChatGPT’s ability to comprehend and apply the Aristotelian Four Causes. The results were promising, demonstrating ChatGPT’s general adeptness in providing illustrations for each cause in relation to specific concepts or entities. Subsequent to this analysis, we investigated the system’s proficiency in crafting follow-up questions anchored in the Four Causes framework. ChatGPT exhibited remarkable efficiency in generating inquiries that were not only relevant to the ongoing conversation but also adeptly integrated aspects from the most recent query, thus maintaining a coherent and contextually relevant learning trajectory. However, it was observed that ChatGPT sometimes struggled to accurately assign the appropriate cause category to each question and to prioritize certain causes over others based on the context or scenario at hand. This suggests a potential area for refinement in the underlying language model, which will be explored more comprehensively in the Discussion section (§ 8.3.3 ). Despite these challenges, the utilization of the generic model facilitated a structured and in-depth exploration of topics, thereby enhancing the self-learning experience by guiding learners through a more nuanced understanding of the subject matter. To further refine the guidance provided and enhance its pedagogical value, we incorporated the persona pattern , as outlined by White et al. (White et al. , 2023 ) . This methodology involves configuring ChatGPT to function as an educational tutor, adapting its way of interacting to more effectively facilitate a learning environment. Within this adapted role, ChatGPT’s responsibilities extend beyond merely answering questions; it is also engineered to proactively generate follow-up questions. For each interaction, the directive “ Provide the top four related follow-up questions based on the previous question using the four causes idea ” is employed. This strategy ensures the elicitation of a set of questions that are not only relevant but also imbued with educational significance, thus fostering a more engaging and productive self-learning experience. 5.3. Usage Scenario This section illustrates how Emily, a Ph.D. candidate with an interdisciplinary background, efficiently self-learned a new knowledge domain (NFTs) with CausaDisco . In her previous self-study endeavors, Emily had often relied on LLM tools to help her gather and consolidate information to build a foundational understanding of new concepts. For example, she mentioned using AskYourPDF 6 6 6 https://askyourpdf.com/zh and Poe 7 7 7 https://poe.com/ to quickly extract information from references. Nevertheless, crafting the right questions can be time-consuming, and although some LLM tools offer follow-up questions, they often lack relevance or do not align with her interests. As a result, these tools typically only provide hints or keywords, leaving her to rely on traditional search engines or to read original materials for in-depth learning. Upon adopting CausaDisco , Emily’s learning process evolved. She first examined the Concept Graph View in the upper right corner to gain an overview of the core concepts. Then, with the help of the navigation bar, she quickly reviewed the Embedded Content View in the upper left corner. As a beginner in NFTs, Emily zeroed in on key concepts and sought clarification through the Q&A Conversation View , asking, “ Can you explain what fungible means in one sentence? ” After reading the reply and realizing that each NFT’s price varies–a topic she found intriguing–she inquired further, “ How do NFT creators determine their products’ prices? ” After receiving this response, she engaged in five turns of interaction with the AI chatbot on the topic. Under the guidance of the follow-up questions, she conducted an in-depth exploration of NFT pricing and the influencing factors, achieving a comprehensive understanding. By reviewing the Tree Map View , Emily noted that she had explored the factors affecting NFT pricing from various angles, including the personal expectations of artists, the volatility of cryptocurrency prices, and the pricing differences between international and local markets. She also understood how to address price volatility. Satisfied with the depth and breadth of her exploration, Emily then started on a new topic—how NFTs influence society—and pursued deep self-learning through a workflow similar to the one she had previously followed. 6. Evaluation We conducted a within-subject study to evaluate whether CausaDisco enhances knowledge sensemaking during self-learning. Specifically, we aimed to investigate the following three research questions (RQs): RQ1: How does CausaDisco foster interactivity with the LLM chatbot? RQ2: How does CausaDisco facilitate sensemaking in self-directed learning? RQ3: How do users perceive the usefulness and experience of interacting with CausaDisco ? 6.1. Participants Our study involved 36 participants (19 male, 17 female) ranging in age from 20 to 36 years old (M = 26.31, SD = 3.01). We recruited participants through convenience and snowball sampling, using criteria similar to those of our formative study. Participants had diverse educational backgrounds, with 13 holding bachelor’s degrees, 15 holding master’s degrees, and 8 holding doctorates. Most participants (n = 25) specialized in STEM fields, while others came from the humanities and social sciences (n = 5), interdisciplinary studies (n = 3) and medical fields (n = 2). Participants also had varying levels of academic training experience (more than five years: n=11; three to five years: n=13; fewer than three years: n=12). Familiarity with LLM chatbots ranged from extremely familiar (n = 3) and very familiar (n = 15) to somewhat familiar (n = 11), somewhat unfamiliar (n = 6) and neutral (n = 1). 6.2. Procedure We employed a within-subject experimental design in which participants independently completed two self-learning tasks, each focused on a distinct concept: NFTs (Schulze, 2024 ) and semiotics (StudySmarter, 2024 ) . To minimize order bias, we counterbalanced the order of exposure to our system, CausaDisco , and a baseline condition (using participants’ preferred tools), resulting in 4 (= 2 x 2) conditions (Guo et al. , 2023 ; Suh et al. , 2023 ) . Participants were encouraged to engage in active learning through conversational interaction with the LLM chatbot in both conditions (see Table 7 in Appendix B ). The study began with participants providing informed consent and completing a pre-study survey with demographic questions. They then engaged in two self-learning tasks, using either CausaDisco or their chosen baseline tools to explore one of the assigned topics. The procedure for self-learning tasks involved four distinct sessions: (1) Introduction: A brief overview of the research background and experimental protocols. (2) Pre-Task Exercise: Participants were given 10 minutes for free exploration of the assigned topic. (3) Self-Learning Tasks: Participants engaged in each self-learning task for 20 minutes using either CausaDisco or their chosen baseline tools. (4) Assessment: Participants engaged in a 15-minute quiz comprising three multiple-choice and three open-ended questions. This quiz, developed in collaboration with domain experts in NFTs and semiotics, assessed the participants’ understanding of the presented concepts. Following the quiz, participants completed a post-task survey using a seven-point Likert scale to evaluate how each condition supported their sensemaking during self-learning. The session concluded with semi-structured interviews where participants compared CausaDisco and the baseline condition, providing feedback on the system design. The entire experiment lasted approximately 90 minutes. The 20-minute duration for each self-learning task was determined based on observations from a pilot study with team members. Figure 7. This figure presents participants’ subjective ratings (seven-point Likert scale) comparing CausaDisco and a baseline on interactivity and sensemaking. 6.3. Measures We evaluated participants’ perceptions of interactivity, sensemaking, and the usability and design of CausaDisco . All measures were rated using a seven-point Likert scale (1 = Strongly Disagree, 7 = Strongly Agree). 6.3.1. Interactivity Measures To assess how CausaDisco enhanced interactivity, we analyzed both objective measures and subjective participant perceptions. We began by analyzing the dialogue records generated during the self-learning tasks, counting and comparing the number of dialogue turns between participants and LLMs in both the CausaDisco and baseline conditions. This analysis provides an objective measure of interaction frequency. To evaluate participants’ subjective experiences of interactivity, we examined two key dimensions: perceived engagement and efficiency (Fig. 7 ). Each dimension comprised two subdimensions. Engagement was assessed by participants’ perceived logical coherence of system-generated follow-up questions ( Q1 ) and their perceived inspiration for exploring new knowledge domains ( Q2 ). Efficiency was evaluated by participants’ perceived ability to efficiently accomplish self-learning objectives ( Q3 ) and the perceived streamlining of the learning process ( Q4 ). 6.3.2. Sensemaking Measures Building upon the sensemaking model proposed by Pirolli et al. (Pirolli and Card, 2005 ) , we assessed how CausaDisco facilitates sensemaking through both subjective and objective measures. For subjective evaluation, we examined two key dimensions: perceived comprehensiveness and sophistication (Fig. 7 ). Comprehensiveness was evaluated by participants’ perceived ability to obtain sufficient information with the system ( Q5 ), their perceived support for comprehensive exploration of new knowledge domains ( Q6 ), and their perceived acquisition of multi-dimensional perspectives during self-learning ( Q7 ). Sophistication was assessed by participants’ perceived deepening of understanding within a new knowledge domain ( Q8 ). For objective evaluation, we recruited domain experts—two Ph.D. students in Fintech and two with master’s degrees in communications—to blindly evaluate participants’ quiz responses. Each participant’s final score was calculated as the average of two expert ratings. The multiple-choice questions assessed the comprehensiveness of understanding, while short-answer questions evaluated the sophistication of knowledge application. 6.3.3. Usability and Design Measures We evaluated how users perceived CausaDisco ’s usability and design using the technology acceptance model (Venkatesh and Bala, 2008 ) . Participants rated their perceptions of the system’s ease of use, ease of learning, and intention to recommend CausaDisco on a seven-point Likert scale. Furthermore, we assessed participants’ perceived intuitiveness of the system, encompassing both interface and interaction design elements. Table 5. Within-subject comparisons of participants’ subjective ratings in the two conditions. Measures include perceived engagement, efficiency, comprehensiveness, and sophistication. Statistical significance: * p < 0.05 p<0.05 , ** p < 0.01 p<0.01 , *** p < 0.001 p<0.001 , * p < 0.1 p<0.1 (marginally significant). Measures Attributes Baseline CausaDisco Statistics Mean/SD Mean/SD t df p Sig Engagement Logically Coherent 5.28/1.28 6.06/0.95 3.08 35 0.001 ** Inspiring 5.31/1.31 6.08/0.65 3.08 35 0.001 ** Efficiency Efficiency Enhancement 5.31/1.04 5.64/1.15 1.83 35 0.038 * Streamlined Process 5.25/1.30 5.53/1.18 1.06 35 0.149 Comprehensiveness Sufficient Information 5.69/1.12 5.69/1.31 0 35 0.500 Comprehensive Exploration 5.31/1.21 5.72/1.32 1.46 35 0.076 * Multi-Dimensional 4.89/1.37 5.67/1.39 2.38 35 0.011 * Sophistication In-Depth Understanding 5.14/1.38 5.78/1.05 2.44 35 0.009 ** 7. Results This section presents findings from our analysis of participant survey responses, prototype interaction logs, and interview feedback. To compare quantitative data from surveys and system logs, we used paired samples t-tests between the two conditions, except when Shapiro-Wilk tests indicated deviations from normality. In those cases, we used the non-parametric Wilcoxon Signed Rank test. 7.1. Interactivity Enhancement (RQ1) To evaluate CausaDisco ’s interactivity, we assessed both perceived engagement and efficiency using a combination of objective and subjective measures (see Table 5 and Fig. 8 ). 7.1.1. Engagement Participants’ ratings of logical coherence are significantly higher in the CausaDisco condition ( M = 6.06 M=6.06 , S ​ D = 0.95 SD=0.95 ) than in the baseline condition ( M = 5.28 M=5.28 , S ​ D = 1.28 SD=1.28 , p < .01 p<.01 ). Similarly, participants’ ratings indicated significantly higher inspiring of the CausaDisco condition ( M = 6.08 M=6.08 , S ​ D = 0.65 SD=0.65 ) than the baseline condition ( M = 5.31 M=5.31 , S ​ D = 1.31 SD=1.31 , p < .01 p<.01 ). Aligned with the self-rating results, participants engaged in significantly more dialogue turns in the CausaDisco ( M = 6.97 M=6.97 , S ​ D = 2.22 SD=2.22 ) than in the baseline condition ( M = 4.08 M=4.08 , S ​ D = 1.86 SD=1.86 , p < .01 p<.01 ). Qualitative data from participant interviews corroborates these quantitative findings. Many participants were impressed by CausaDisco ’s epistemologically informed follow-up questions, describing them as “ more logically structured, providing a clear train of thought ” (P35) and “ more human-like with high-quality ” (P34). Furthermore, participants reported that CausaDisco facilitated the discovery of new points of interest. For instance, P25 stated, “ The follow-up questions and concept graph of CausaDisco are eye-openers! It is like having a curiosity engine–the more I use it, the more I want to keep chatting with the AI. ” The combination of quantitative and qualitative results provides compelling evidence for CausaDisco ’s effectiveness in enhancing user engagement with LLM chatbots. 7.1.2. Efficiency Participants reported significantly higher efficiency enhancement in the system condition ( M = 5.64 , S ​ D = 1.15 M=5.64,SD=1.15 ) compared to the baseline condition ( M = 5.31 , S ​ D = 1.04 , p < .05 M=5.31,SD=1.04,p<.05 ). Participants in the system condition, on average, rated the streamlined process 5.53 ( S ​ D = 1.18 SD=1.18 ), higher than the baseline condition ( M = 5.25 , S ​ D = 1.30 M=5.25,SD=1.30 ), but there are no significant differences. The results underscore the potential of our system to enhance user efficiency in a significant yet implicit manner. Although participants reported perceiving an increase in their efficiency, this improvement was not attributed to a more streamlined process. As P4 remarked, “ The follow-up questions of CausaDisco are spot-on and boost my self-learning efficiency. I learn faster without having to rack my brain thinking up questions myself. It is a real time-saver. ” This suggests that our system may boost participants’ efficiency by delivering precisely the questions they require at the moment. However, the underlying logic and mechanism responsible for generating the follow-up questions are not fully comprehensible to participants. For instance, P29 noted, “ I often pause to consider the underlying logic of the follow-up questions and the learning trajectory in the tree map. The deeper I go with the questions, the more I find myself taking these little moments to reflect. ” This uncertainty may lead participants to doubt the effectiveness of the questions in streamlining their self-learning. This hypothesis is corroborated by the lack of significant findings in streamlined process, for future research and system improvements to better support users’ self-learning. Figure 8. Within-subject comparison of participants’ subjective ratings of key measures (engagement, efficiency, comprehensiveness, and sophistication) across two conditions. Error bars represent between-subjects standard error ; asterisks indicate statistical significance from paired t-tests (*p ¡ 0.05, **p ¡ 0.01). The difference in perceived comprehensive exploration, though not statistically significant, approached significance and is highlighted in gray. 7.2. Sensemaking Support (RQ2) To assess how CausaDisco facilitates sensemaking, we compared participants’ perceived comprehensiveness and sophistication when processing information under the two conditions (see Table 5 and Fig. 8 ). 7.2.1. Comprehensiveness We found no significant differences between conditions regarding sufficient information (the system condition: M = 5.69 , S ​ D = 1.31 M=5.69,SD=1.31 ; the baseline condition: M = 5.69 , S ​ D = 1.12 ; p > .05 M=5.69,SD=1.12;p>.05 ). However, participants reported marginally significantly higher comprehensive exploration in the system condition ( M = 5.72 , S ​ D = 1.32 M=5.72,SD=1.32 ) compared to the baseline condition ( M = 5.31 , S ​ D = 1.21 , .05 < p < .10 M=5.31,SD=1.21,.05 .05 p>.05 ). The results demonstrate that our system did successfully deliver multi-dimensional information to users. However, the scope of the information was constrained by the specific study materials provided. Consequently, participants expressed mixed opinions on whether the new information was sufficient for a fully comprehensive understanding. In line with our goal of offering diverse perspectives on unfamiliar topics, P30 shared, “ Compared to my usual tools, CausaDisco’s follow-up questions focus on key ideas while encouraging broader thinking. Along with the concept map, which gives me a big-picture view of concept connections, CausaDisco helps me grasp new information in a more well-rounded way. ” This achievement also aligns with our primary objective of providing a comprehensive self-learning environment. However, we encountered limitations in conclusively proving that our system offers sufficient information. For instance, P9 commented, “ It seems all info CausaDisco gives, including the follow-up questions and responses, comes straight from the study materials. That’s good for accuracy, but adding some outside facts with sources would help my learning even more. ” This perceived lack of information likely arises from the contrasting exploration approaches between the two conditions. CausaDisco guides users through a specific article, generating precise, content-based follow-up questions and responses for accuracy. Conversely, the baseline condition allows unrestricted exploration. This key difference– CausaDisco ’s focused, article-centric approach versus the baseline’s open-ended exploration–likely created an impression of limited information within our system. 7.2.2. Sophistication Participants reported significantly higher in-depth understanding ratings in the system condition ( M = 5.78 M=5.78 , S ​ D = 1.05 SD=1.05 ) compared to the baseline condition ( M = 5.14 M=5.14 , S ​ D = 1.38 SD=1.38 , p < .01 p<.01 ). Subjective quiz scores were also higher in the system condition ( M = 5.06 M=5.06 , S ​ D = 0.98 SD=0.98 ) than in the baseline condition ( M = 4.56 M=4.56 , S ​ D = 1.12 SD=1.12 , .05 < p < .1 .05