ABSTRACT. Generative AI writing feedback systems provide immediate support for EFL learners, yet concerns remain regarding cognitive offloading, where learners delegate cognitive work to AI rather than actively engaging their own strategies. The study examined how 111 Japanese university students engaged with an AI writing feedback system over a 14-week English writing course. To assess changes in self-regulated learning and cognitive strategy application, engagement patterns were analyzed longitudinally at two levels: across six writing topics to track adaptation to the system, and within each topic to trace error correction trajectories across four linguistic categories. Results showed significant improvement in writing performance, but an overall decline in cognitive strategy use. Patterns diverged markedly by learning outcome groups. High-outcome students sustained cognitive strategy use and engaged in deep-level processing of syntactic and lexical errors, progressively reducing newly introduced errors. In contrast, low-outcome students showed narrowing engagement focused on surface-level corrections and increasing passive dependence on AI feedback. These findings suggest that AI writing feedback effectiveness depends critically on learners' regulatory engagement patterns, with prolonged, excessive cognitive offloading among less successful learners being associated with a decline in their self-reported general cognitive strategy use.

ABSTRACT. Pre-writing is the initial step in writing a persuasive English paragraph. However, this task proves challenging for beginners of English paragraph writing. Therefore, in the previous research, a support prototype system for organizing an English paragraph outline was developed and utilized in actual classrooms. During the classroom practice, some students were found to be stuck in the idea organization phase. To address this problem, we intend to implement a new feature called Intent-Adaptive Support into the support system. Realizing this feature requires three functions, which are "operational state recognition," "intent inference," and "adaptive support." As a first step toward implementing Intent-Adaptive Support, this research aims to develop the function of operational state recognition. This function is designed to track the user's progress. Specifically, it recognizes the user’s operational state by monitoring the user’s interactions. Our evaluation showed that applying operational state recognition to user log data enables the system to understand the user’s situation.

ABSTRACT. Automated writing evaluation systems commonly provide textual feedback for second language writing, but such feedback often focuses learners’ attention on surface-level errors rather than content development. Drawing on the Noticing Hypothesis, this study examines whether AI-generated images can serve as an alternative feedback modality that redirects attention toward meaning construction. It investigates how image-based feedback influences learners’ interactions with an AI-assisted writing system and whether it supports content development during revision. We developed Avery, an AI-assisted writing support system that provides either textual or image-based feedback on descriptive writing tasks. Sixty-seven Japanese university students were randomly assigned to an image-feedback group or a text-feedback group. Participants completed writing tasks, revised their work after receiving feedback, and resubmitted their texts. Interaction logs and revision records were analyzed using established revision frameworks and Transition Network Analysis. Analysis of 95 writing sessions and 1,630 interaction events showed that the image-feedback group exhibited a denser transition network, indicating more varied interaction patterns. These learners more frequently progressed to subsequent writing tasks, viewed feedback, and checked the leaderboard, while showing fewer behaviors associated with leaving and returning to the system. This suggests higher engagement and sustained participation. Revision analysis revealed similar rates of content-level revisions across groups, but substantially fewer surface-level corrections in the image-feedback condition. While textual feedback encouraged linguistic correction, image-based feedback promoted engagement, task continuation, and attention to meaning. These findings highlight the potential of AI-generated images as an alternative feedback modality for supporting L2 writing development.

ABSTRACT. MMLA systems estimate educational events from sensor data, but current LRS designs record uncertain estimates as definitive facts, risking distortion of instructional decisions. This study proposes a five-layer hierarchical architecture (Raw Data, Feature, Interpretation, Event, Context) that preserves inference uncertainty and controls its propagation to educational decision-making. We evaluate the architecture using a teaching behavior recognition system and confirm that the Event layer gate mechanism effectively suppresses pedagogically harmful misclassifications while minimizing loss of correct events. We discuss implications for interoperability and cross-institutional data exchange.

ABSTRACT. While multiple-document integration is vital for knowledge construction, how readers dynamically allocate cognitive resources during digital reading remains under- explored. This study aims to combine eye-tracking technology with machine learning algorithms to identify key reading behaviors that predict integration performance. The primary goal is to provide an empirical foundation for developing real-time feedback and navigation support systems in digital learning environments. The dataset consisted of 122 observations from 62 participants, each of whom completed two reading topics under a consistent task instruction (either summary or argumentation). For each topic, participants read four conflicting texts interconnected with hyperlinks, with each text structured into three functional paragraphs: source information, evidence, and conclusion. To capture the complexity of the reading process, we extracted indicators including first-pass reading time and rereading time for each paragraph, saccade counts between paragraphs, and hyperlink clicks. Data analysis first utilized K-means clustering to categorize readers into high- and low-integration groups based on three writing indicators: concept coverage, information transformation, and the use of connective words. Subsequently, Decision Tree and XGBoost algorithms were applied to predict group membership based on process data and task instruction. The results revealed that evidence rereading time and task instruction were the most critical predictors of high-quality integration. Specifically, while the summary instruction generally increased the probability of high-integration classification, the decision tree revealed that this advantage was conditional: readers under the summary instruction still fell into the low-integration group if they demonstrated insufficient initial processing or low navigation engagement. Conversely, sufficient rereading time on evidence paragraphs emerged as a prerequisite for success across both tasks. These findings suggest that digital reading platforms should leverage real-time behavioral data and task demands to provide personalized scaffolding, thereby optimizing learners' performance in multiple-document integration.

ABSTRACT. This study examines the role of large language model (LLM)–powered tools in supporting adult learners in the Philippine Alternative Learning System (ALS) during the development of their Portfolio of Evidence (PoE). Grounded in Self-Determination Theory (SDT), the study investigates how LLM-assisted writing environments may influence learners’ autonomy, technical competence, relatedness, confidence, and motivation in non-formal education settings. A sequential explanatory mixed-methods design is employed, consisting of quantitative and qualitative phases. The quantitative phase utilizes a structured survey questionnaire analyzed using Partial Least Squares Structural Equation Modeling (PLS-SEM), while the qualitative phase uses semi-structured interviews to explore learners’ experiences with LLM-assisted PoE writing. Prior to the main implementation, pilot testing was conducted among 30 ALS adult learners with prior experience using LLM-powered tools such as ChatGPT, Gemini, Microsoft Copilot, Claude, and Grammarly. Preliminary pilot testing results demonstrated acceptable reliability and validity. Cronbach’s Alpha and Composite Reliability values exceeded the recommended threshold of 0.70, while Average Variance Extracted (AVE) values surpassed the acceptable level of 0.50. Discriminant validity was established using the Heterotrait-Monotrait Ratio (HTMT) and Fornell-Larcker Criterion. Low-performing indicators were removed to refine the survey instrument for full-scale implementation. The study contributes to emerging discussions on AI-assisted learning in non-formal education and provides preliminary evidence supporting the use of SDT as a framework for examining LLM-assisted writing among ALS adult learners.

ABSTRACT. Adoption of learning analytics is hindered by a lack of locally relevant practice cases. To bridge this gap, we developed the LEAF Evidence Portal, a repository of case articles that describe the adoption of learning analytics in classrooms. The case articles are tagged by different tag categories such as school level and tool. In this paper, we analyze two months of portal access logs to identify what drives user exploration of case articles. The results show that learning analytics tool is the primary driver for user interest and co-browsing. Conversely, users tend to browse across different subjects. These findings suggest that recommender systems for case articles should prioritize tools over subject-specific filtering to better support learning analytics adoption.

ABSTRACT. With the increasing integration of Generative Artificial Intelligence (GenAI) into K-12 education, the initial level of primary school students' prompt engineering remains insufficiently explored through systematic empirical research. Using a participatory observation methodology, this article analyzes primary school students' self-directed human-AI interaction processes in the context of learning tasks from both behavioral and cognitive perspectives, examining their prompt engineering and exploring the relationship between different modes of student-generated prompt and the quality of AI-generated outputs. This article reveal that primary school students have already possessed a foundational level of prompt engineering and demonstrate distinctive cognitive and behavioral characteristics. Crucially, students who express prompts in a written style mode tend to obtain higher-quality generated outputs than those using spoken-style modes, which better supports the completion of learning tasks. Finally, attempting to break through generic prompt frameworks,this article proposes a two-dimensional “elements–features” framework based on the existing “starting point” of primary school students' prompt engineering. This framework provides theoretical and practical references for enhancing students' prompt engineering and for promoting GenAI applications from a student-centered perspective in K-12 education.

ABSTRACT. While generative AI offers highly adaptive scaffolding in education, its pervasive assistance risks undermining learner autonomy. Existing research predominantly treats LLMs models as the option of technical accuracy rather than analyzing how distinct model characteristics impact educational effects. To address this gap, this study proposes a paradigm shift toward active, self-regulated GenAI utilization. We integrated three distinct generative-AI models into an AI-powered math scaffolding tool, having students actively evaluate and select the model that best fits their learning needs. To evaluate this approach, we conducted an empirical study utilizing an adaptive question-generation tool and an e-book platform. We employed a Difference-in-Differences framework and an ordinal logistic regression model to rigorously analyze causality and the underlying psychological mechanisms. The results revealed distinct pedagogical trade-offs among the models regarding descriptive volume, conceptual density, and generation latency. Crucially, the sheer act of autonomously selecting a model specifically enhanced students’ subjective understanding. This selection did not artificially inflate motivation or relevance, rather than that these metrics remained grounded in students' prior academic abilities and system latency. Ultimately, this study demonstrates that only one LLM model would not be selected in technology design. Embedding explicit opportunities for autonomous model selection can empower human agency.

ABSTRACT. To address the marginalization of the Information Technology (IT) discipline and superficial knowledge integration in current high school interdisciplinary instruction, this study constructs and empirically validates a competence-oriented, five-stage implementation framework encompassing theme selection, conceptual linkage and problem chain organization, competence-based goals, 5E inquiry processes, and comprehensive assessment. Through a one-semester, three-cycle action research involving 52 tenth-grade students exploring the theme "My 'Superpower' Persona," mixed-methods data demonstrated that the framework significantly enhanced students' computational thinking, information awareness, digital learning and innovation, and interdisciplinary competencies. Furthermore, classroom engagement and artifact quality improved substantially. Although the growth in information social responsibility was non-significant—indicating the need for longer-term cultivation—the proposed framework effectively safeguards the core educational value of the IT discipline. It successfully overcomes the "knowledge patchwork" dilemma, providing a replicable and systematic paradigm for high school IT interdisciplinary instruction.

ABSTRACT. Automated Short Answer Grading (ASAG) leveraging Large Language Models (LLMs) offers significant potential for reducing teacher workload in educational assessment. However, applying these systems to morphologically complex, low-resource languages like Thai presents unique challenges. A critical but underexplored limitation is "keyword dependency" the tendency of Artificial Intelligence (AI) to evaluate responses based on surface-level lexical matching rather than genuine semantic understanding, creating a grading inequity where students who paraphrase correct answers may be penalized, while those who use expected vocabulary without understanding may receive inflated scores. This study investigates the robustness of five major LLMs, namely ChatGPT-4o, Gemini 2.0 Pro, Claude 4.0 Sonnet, Grok 4, and Typhoon 2 Instruct (70B), against adversarial student response conditions in grading open-ended Thai-language written exams. As a core methodological contribution, we introduce a controlled adversarial testing design using six behavioral profiles: four representing normal student performance levels, and two adversarial cases, a Keyword Adversarial profile (responses containing expected keywords but with intentionally distorted semantics) and a Semantic Adversarial profile (semantically correct responses that deliberately avoid expected rubric keywords). All models were evaluated against a baseline of two professional Thai teachers using a 55-question dataset, assessed through a comprehensive 5-metric evaluation framework covering both numerical score agreement and explanatory faithfulness via reason codes. Under normal student response conditions (Students 1–4), the LLMs achieved MAE ranging from 0.218 to 0.400. Under the Keyword Adversarial condition (Student 5), the models exhibited clear keyword inflation, overscoring semantically flawed answers simply because they contained expected vocabulary. Most critically, under the Semantic Adversarial condition (Student 6), the models demonstrated extreme semantic undervaluation, systematically penalizing correct answers that utilized paraphrasing or synonyms, with Weighted Cohen's Kappa collapsing to as low as 0.067 and Krippendorff's Alpha falling well below acceptable reliability thresholds at 0.422. Reason code analysis directly confirmed the AI's structural reliance on lexical cues over logical correctness, revealing that current AI systems penalize keyword absence more severely than semantic incorrectness, a finding with direct implications for grading fairness in synonym-rich languages like Thai.

ABSTRACT. Lesson plan alignment, the consistency of cognitive depth between teaching objectives and instructional procedures, is critical for developing higher-order thinking. However, existing lesson plans often suffer from cognitive misalignment, wherein high-standard objectives are reduced to low-level tasks in practice. This study employs computational text analysis to compare cognitive alignment between human-authored lesson plans (HLPs) and Multi-Agent System (MAS) generated plans. Using a corpus of 288 instructional segments, targeted dictionary matching mapped instructional verbs to Webb’s Depth of Knowledge (DOK) framework, while Structural Topic Modeling (STM) identified macro-level pedagogical distributions. The results reveal two key findings. First, at the micro-lexical level, HLPs exhibit significant cognitive misalignment, heavily relying on DOK Level 1-2 verbs despite setting higher-order objectives. Conversely, MAS-generated plans demonstrate a commanding presence of Level 3-4 analytical verbs. Second, at the macro-structural level, STM analysis uncovers a critical pattern of content convergence alongside pedagogical divergence. While HLPs and MAS align perfectly in factual content coverage, MAS significantly outperforms HLPs in constructing explicit higher-order instructional scaffolds, effectively avoiding the pervasive “activity trap.” Ultimately, this study redefines cognitive misalignment as a procedural translation gap and proposes a data-grounded workflow for human-AI collaborative instructional design. It provides empirical evidence that while human educators are essential for contextual anchoring, the MAS functions as a vital structural enforcer, empowering teachers to design rigorous learning experiences without weakening pedagogical alignment.

ABSTRACT. With the advancement of educational reform, instructional support has become a critical indicator for evaluating teaching effectiveness. However, existing evaluation frameworks for instructional support remain largely subjective and difficult to quantify. To address these limitations, this study integrates multimodal classroom data and artificial intelligence (AI) techniques to develop a multimodal evaluation framework for classroom instructional support. Grounded in the Classroom Assessment Scoring System (CLASS) framework and established educational theories, we developed an initial evaluation framework through a systematic literature review. We then refined and optimized the framework through expert interviews and a two-round Delphi method, and quantified indicator weights at all levels using the Analytic Hierarchy Process (AHP). A complete framework with 3 primary indicators and 7 secondary indicators was finally established. We verified the framework’s validity using 30 classroom videos analyzed by a customized Generative Pre-trained Transformers (GPTs) annotation model, which supports a zero-code automated annotation workflow. This study offers a theoretical basis and practical tool for the standardized evaluation and precise diagnosis of multimodal classroom instructional support.

ABSTRACT. Mathematics learners often develop procedural competence without corresponding conceptual understanding, in part due to limited opportunities for structured reflection in classroom-oriented digital environments. Prior research shows that reflection and self-explanation can support learning, but their design must balance effectiveness with instructional efficiency in real-world settings. We investigate the impact of Large Language Model (LLM)–supported reflection prompts embedded in a digital mathematics learning game for 5th- and 6th-grade students across two classroom studies. In Study 1 (N = 191), we compared three reflection formats following problem solving: (1) multi-turn, LLM-personalized Socratic dialogue, (2) open-ended self-explanation, and (3) multiple-choice self-explanation. All conditions produced significant learning gains, with no differences between formats; however, the more generative formats required substantially more time (1 and 2), indicating increased interaction burden without added benefit. To isolate whether these null effects were due to extended interaction or personalization, Study 2 (N = 33) compared a single LLM-personalized Socratic prompt with a standard open-ended self-explanation. Results again showed comparable learning gains, with no differences in outcomes or duration. Together, these findings identify a boundary condition for LLM-supported reflection: increasing dialogic interaction or personalization does not inherently enhance learning for 5th- and 6th-grade students and may introduce unnecessary interaction costs. For technology-enhanced learning systems, these results provide actionable design guidance -- prioritizing concise, targeted reflection over extended dialogue may better align with young students in the classroom. We discuss implications for designing efficient, classroom-ready reflective supports and consider when dialogic approaches may be more appropriate, such as for older learners.

ABSTRACT. Accurate information retrieval from learning materials is a prerequisite for trustworthiness in many Artificial Intelligence in Education (AIED) applications, including Retrieval-Augmented Generation (RAG) and personalized learning frameworks. However, the effectiveness of general retrieval models in capturing complex pedagogical relationships remains largely unexplored, hindering the selection of effective models for real-world learning environments. This paper presents a technical contribution to this educational and domain-specific retrieval by introducing a benchmark of 1,470 annotated slide pairs designed to evaluate the retrieval of pedagogically meaningful relationships. In addition, we propose two methods to adapt language models to this educational setting. Our findings provide practical guidance for model selection, showing that OpenAI’s embeddings perform strongly among general models and demonstrate that our adaptation methodologies lead to an improvement of retrieval by up to 43% for RoBERTa and 6% for OpenAI embeddings (PR-AUC). Finally, our approach enhances interpretability by providing clearer retrieval rationales, supporting transparency and trust in AI-driven educational systems.

ABSTRACT. In creative activities, it is crucial for learners to actively regulate the cyclical process of SRL (forethought, performance, and self-reflection). This study treats daily trial-and-error in research activities as "learning materials" and proposes a support system to foster better self-regulated learners. Specifically, we developed a framework that structures SRL in research into three layers: (1) problem-solving activities (RA layer), (2) SRL activities (SRLA layer) for process control, and (3) meta-learning activities (MLA layer) for acquiring metacognitive knowledge from the process. The proposed system features functions to organize thought products into a chain of "questions" and "answers" (RA layer), externalize solution plans as a hierarchical structure of "goals and means" (SRLA layer), and structurally support activities to reconstruct "lesson learnt" (metacognitive knowledge) based on these processes (MLA layer). A practical implementation involving 14 students confirmed that the three-layered processes can be continuously executed and externalized within daily research activities. The results suggested that chronological activity logs functioned as "learning resources" that drive dialogue with supervisors, facilitating the refinement of specific metacognitive knowledge tailored to individual cognitive characteristics.

ABSTRACT. Nonviolent Communication (NVC) offers a structured framework for empathetic conflict resolution, yet internalizing its principles into spontaneous behavior remains a formidable challenge for adolescents. This study examined whether an "Implicit Adaptive Matching" mechanism — in which a custom AI dialogue system covertly diagnoses a learner's immediate psychological need (emotional support vs. problem-solving) and routes them to a matched AI persona (Warm vs. Competent) — produces superior NVC learning outcomes compared to a standardized baseline AI. Using a nonequivalent pretest-posttest quasi-experimental design, 86 valid samples of eighth-grade students were retained for final analysis after a two-week intervention. A manipulation check confirmed that students perceived the intended AI persona differences (t(54) = 2.754, p = .008). An ANCOVA controlling for pre-test scores revealed a marginally significant advantage for the Adaptive AI group on the Nonviolent Communication Behavior Scale (NVCBS) post-test (F(1,83) = 3.632, p = .060, ηp² = .042). Qualitative analysis of chat logs identified the "Jackal language cognitive loop" as the core learning barrier, and demonstrated how the adaptive mechanism provided precisely tailored affective and cognitive scaffolding to overcome it. System Usability Scale (SUS) scores confirmed that interface usability was uniformly high across all conditions (F(2,108) = 0.645, p = .527), ruling out technology as a confound. These findings advance knowledge of AI design grounded in the Stereotype Content Model (SCM) for social-emotional learning (SEL) and provide a replicable blueprint for adaptive affective computing in middle school counseling contexts.

ABSTRACT. On-demand video lectures frequently induce a passive learning state, leading to a loss of concentration and an "illusion of competence". Furthermore, a significant problem exists in that learners rarely take the initiative to seek help—such as by conducting independent research or asking questions—even when they encounter difficulties during the learning process. To solve this problem, interactive systems utilizing virtual agents and Large Language Models have been proposed. However, they lack the physical embodiment often considered necessary to assist in re-engaging distracted learners. Furthermore, these digital systems struggle to chronologically synchronize the LLM's knowledge base with the learner's real-time progress through the video material. To address these challenges, this study proposes a novel active learning support system that integrates a physical social robot (Kebbi) with a "Time-Adaptive Retrieval-Augmented Generation" approach. The system automatically pauses the video at educationally significant moments and proactively intervenes using the robot's physical gestures (e.g., head tilting) to elicit spontaneous questions. Furthermore, Time-Adaptive RAG dynamically restricts the LLM's reference knowledge to the transcript of the video watched up to that exact moment, enabling contextually synchronized dialogue. A preliminary feasibility study with eight university students was conducted to evaluate the system. The quantitative and qualitative results provided initial support for the feasibility and acceptability of the proposed system. First, the dialogue logs suggested that Time-Adaptive RAG generated contextually appropriate responses aligned with the learner’s current video progress (H1). Second, the proactive interventions appeared to encourage spontaneous questioning, while head pose data showed no significant increase in postural collapse (H2). Third, the combination of automatic video pausing, the robot's subtle physical affordances, and seamless progression control based on LLM intention estimation was generally accepted by participants without obvious signs of fatigue or visual deviation (H3). In conclusion, integrating a physical robot with temporally synchronized LLM dialogue suggests a feasible and acceptable approach to transforming passive video viewing into an engaging, active learning experience.

ABSTRACT. Large language models (LLMs) are increasingly entering students’ learning practices, but their educational value depends on whether they are used to support reasoning or to complete tasks without engaging in the underlying reasoning. This study examines guided LLM use in an undergraduate Probability and Statistics course, focusing on the distinction between assigned LLM access and the quality of students’ actual interaction with the model. In a four-week quasi-experimental summer program, students were organized into three balanced conditions: no LLM access, unrestricted LLM access, and guided LLM access. The guided condition used the same LLM platform as the unrestricted condition, but students received explicit training and rules intended to promote reasoning-focused help-seeking, stepwise hints, verification, and ethical use. All quizzes and the delayed final exam were completed without LLM or external assistance, allowing us to separate AI-supported practice performance from independent learning. Results show that guided use was associated with a clearer learning-oriented interaction pattern than unrestricted access, especially in prioritizing reasoning over final answers and requesting stepwise support. Guided-LLM students showed a promising pattern of stronger no-help quiz performance in the intervention phase, while unrestricted access appeared more useful for assisted practice completion than for consistently improving independent performance. Available time measures did not support a simple duration-based explanation, and self-assessment calibration suggested better alignment between perceived and demonstrated understanding in Guided-LLM. Overall, the findings suggest that LLM access alone is an incomplete educational intervention. For Artificial Intelligence in Education (AIED), the central design challenge is to scaffold how students use LLMs so that these systems function as partners in reasoning rather than answer-getting tools.

ABSTRACT. AI agents are increasingly deployed as interactive partners in immersive learning environments (ILEs), yet existing reviews provide limited synthesis of the educational roles these agents perform, the techniques that enable them, and the challenges reported in empirical implementation. This paper reports selected findings from a systematic review of 22 studies, guided by two research questions on agent types/techniques and implementation challenges. Across the reviewed studies, AI agents serve four educational roles: scaffolding and instructional mediation, orchestration and process adaptation, assessment and diagnosis, and companion and persona. Their technical implementations have shifted toward multimodal hybrid architectures integrated with ILE platforms, combining generative language and conversational pipelines, learner modelling and adaptive control, structured knowledge and rule-based reasoning, and embodied multimodal integration. Reported challenges are clustered into pedagogical issues, including support versus learner agency, personalisation versus appropriateness of support and presence versus instructional coherence, as well as infrastructural constraints, including interaction latency, physical and attentional load, and complex real-time system integration. Overall, AI agents in ILEs are best understood as contributors and collaborators in a learning process rather than as fixed platform features. The review also suggests aligning agent roles with task phases and learner proficiency, reporting evidence of interaction processes alongside learning outcomes, and exploring emerging capabilities such as agent memory, skills, and autonomy.

ABSTRACT. AI tutors can appear successful under proxy objectives while failing to support learning. A response may be fluent, confident, or correctly formatted, yet still skip reasoning, accept a misconception, or provide invalid intermediate steps. This paper proposes an evaluation-time diagnostic framework for identifying pedagogi- cal reward-hacking risks in AI tutor outputs. The framework combines adversarial prompt stress tests, automatic behavioral detectors, human pedagogical ratings, and multi-objective scoring. Using Claude Haiku 4.5 as the common foundation model, we evaluate five prompt-level conditions on a balanced GSM8K sample: base, reward shaping, RLHF-style preference prompting, CPO-style constraints, and the proposed diagnostic condition. No model weights are updated, so the study should be read as diagnostic evaluation rather than training-time mitiga- tion. Results show that the proposed condition improves over prompt-simulated diagnostic baselines on final-answer accuracy and receives the strongest human pedagogical ratings, while base Claude remains strongest on raw answer accuracy.

ABSTRACT. Predicting student dropout is a pivotal concern for educational institutions, as timely identification of at-risk students enables them to implement early interventions aimed at reducing attrition. This paper proposes an ensemble deep learning model that combines the application of convolutional (CNN) and long short-term memory (LSTM) networks to analyze both spatial feature patterns and temporal dependencies in student academic data. To address the class imbalance inherent in dropout datasets, we apply the SMOTETomek Link technique, which improves the model's sensitivity to the minority class. The proposed CNN-LSTM ensemble aggregates the outputs of five independently trained models to enhance generalization and reduce overfitting. Our model supports both tabular and time-series input formats by transforming student semester-wise data into sequential representations. Experimental evaluation using a real-world dataset from the Polytechnic Institute of Portalegre demonstrates that the ensemble CNN-LSTM model outperforms independent CNN and LSTM model architectures. The best-performing configuration demonstrated an accuracy of 90.38%, a recall of 91.83%, and an F1-score of 90.31%. These results highlight the efficacy of the proposed approach for robust and early dropout prediction, with strong potential to support data-driven decision-making in educational institutions.

ABSTRACT. Blended learning has become a prominent modality in higher education due to its high degree of spatiotemporal flexibility and autonomy. However, spatial transitions and highly autonomous environments pose challenges in maintaining high levels of student engagement, and traditional static evaluation metrics struggle to capture the dynamic evolutionary patterns of engagement during the learning process. Therefore, this study innovatively introduces Dynamic Systems Theory (DST) to conduct an intensive longitudinal tracking of the multidimensional engagement trajectories of 30 undergraduate students over an 8-week (16 observations) blended learning course. Utilizing analytical methods such as moving extreme value plots and moving correlation coefficient plots, the study reveals that: (1) student engagement in blended learning environments exhibits nonlinear development. Specifically, dimensions like cognitive engagement align with the U-shaped development pattern in DST, and external environmental transitions are prone to trigger unstable systemic oscillations. (2) Complex self-organizing relationships exist among the behavioral, emotional, and cognitive sub-dimensions, manifesting in three typical development models: "behavioral-supported", "emotional-competitive", and "fully-supported". This study provides a novel dynamic perspective for understanding student learning mechanisms in blended learning and offers empirical support for implementing targeted pedagogical interventions.

ABSTRACT. This study examined how gender-related differential item functioning (DIF) in a performance-based digital literacy assessment can be interpreted using log data. Log data from two items flagged for gender DIF in the 2024 National Digital Literacy Assessment were analyzed: Item 2-8 (a functional task) and Item 3-5 (a production task). For Item 2-8, behavioral sequence data were constructed and analyzed using optimal matching and agglomerative hierarchical clustering. For Item 3-5, time- and action-based variables were derived and analyzed using K-medoids clustering with Manhattan distance. The results showed that the identified behavioral clusters were associated with both performance outcomes and gender distributions. For Item 2-8, the DIF favoring males was associated with differences in file-selection strategies and program execution processes. In Item 3-5, the DIF favoring females was associated with differences in checking and applying the required conditions. These findings suggest that log data can provide process-based evidence for interpreting DIF and examining the fairness and validity of performance-based assessments.

ABSTRACT. Real-time prediction of student failure may enable timely intervention during programming activities. Using 29,059 Java quiz attempts from 877 students, we investigate whether eventual failure can be predicted from partially observed interaction traces. We analyze 121 debugging and code-writing questions using a prefix-based framework and compare a Hidden Markov Model (HMM), a Long Short-Term Memory network (LSTM), and XGBoost. At a representative threshold of 0.70, HMM and XGBoost achieve higher recall and longer median lead times than LSTM, while LSTM produces substantially lower false-positive rates. Median lead times range from approximately 5–10 minutes across models and task types, indicating that partial interaction traces can provide an actionable window for intervention. These findings demonstrate the feasibility of early-risk detection from programming interaction logs while revealing important tradeoffs among detection coverage, alert timeliness, and false-alarm rates.

ABSTRACT. Measuring online learning self-efficacy is important for understanding learner success in today’s ubiquitous digital environments. A wide variety of measurements have been developed over the decades, yet the reliability of these assessments is often assumed rather than empirically verified, and their use remains scattered. Adhering to the PRISMA statement (Moher et al., 2009; Page et al., 2021) and the REGEMA checklist (Sánchez-Meca et al., 2021), this reliability generalisation (RG) meta-analysis synthesised Cronbach’s alpha coefficients from studies using three prominent online learning self-efficacy measures: Miltiadou and Yu (2000), Tsai et al. (2020), and Zimmerman and Kulikowich (2016). We screened 20,318 records from Scopus and Web of Science (2000-2025), yielding only 15 eligible articles reporting usable reliability data (k = 18, N = 4,581). The pooled overall reliability was excellent (α = .9395, 95% CI [.9156, .9566]). However, extreme heterogeneity was also observed (I² = 97.54%). Scale-specific estimates ranged from α = .8395 (Tsai et al.) to α = .9775 (Zimmerman & Kulikowich), with substantial between-study variance (τ² up to 0.8283). Moderator analyses revealed that reliability was not an invariant property. In particular, gender composition, geographical region, Likert format, scale modifications, and reporting of factor analyses significantly influenced estimates. Importantly, publication bias tests suggested no systematic suppression of low reliability studies. However, as an important finding, this meta-analysis is constrained by a very small number of primary studies (k = 18 across all scales), limiting statistical power for moderator analyses and precluding meta-regression with multiple simultaneous covariates. Most eligible studies failed to report McDonald’s omega or test-retest reliability, and few employed probability sampling. Hence, we suggest researchers should (a) routinely report score reliability for their own sample rather than citing past values, (b) conduct and report confirmatory factor analysis, (c) include diverse samples beyond convenience-based university students, (d) choose suitable and validated measures for particular samples with full explanations provided.

ABSTRACT. This pilot study explores how physiological and self-report measures can be combined to assess emotional engagement in immersive learning. Seven university students participated in a within-subject experiment comparing head-mounted display (HMD)-based learning materials with slide-based materials in the context of a multinational Virtual Exchange (VE). Heart rate and skin conductance were measured, along with self-reported emotional states using the General Affect Scale. The analysis revealed significant interactions between the learning format and the measurement period for both skin conductance and for positive and negative affect. Specifically, the HMD-based materials were shown to elicit higher emotional arousal during the task compared to the slide-based materials. These preliminary findings suggest the utility of an integrated approach, combining physiological signals with traditional measures, to deepen the understanding of emotional engagement in XR-based education.

ABSTRACT. Field-based learning, such as fieldwork, plays a vital role in bridging the gap between theory and practice. Specifically, reflection is essential for enabling learners to restructure their understanding of prior classroom learning. However, learners often struggle to relate their practical experiences to theoretical knowledge. Despite advancements in technology-enhanced learning approaches, few studies have integrated both field-based and classroom learning contexts to address this difficulty. To fill this gap, this study proposes a context-aware learning environment called AllerColler. It is designed to provide personalized feedback that facilitates sense-making across both learning settings during reflection. The proposed system leverages a large language model (LLM) to estimate relevant classroom learning content based on learner inputs. To ensure the feasibility of the proposed system in practical educational settings, semi-structured interviews were conducted with five experts in field-based learning in Japan. A thematic analysis revealed that the proposed system could navigate field-based learning where the learning objectives are predetermined or guided by educators. Additionally, the content of personalized feedback could facilitate reflection, although a potential risk of overreliance on intelligent support was identified. These findings clarify future directions for this approach to promote the thoughtful mobility of learning across classrooms and real-world settings.

ABSTRACT. Online learning faces persistent challenges in sustaining student engagement, and while adaptive gamification has been proposed to address the inconsistent effects of one-size-fits-all gamification approaches, most implementations rely on static adaptation that overlooks the temporal evolution of learner motivation. This paper presents the design, implementation, and preliminary evaluation of a three-stage Dynamic Gamification Adaptation (DGA) algorithm. The algorithm establishes player types via the validated Hexad scale, maps each player type to a set of game elements grounded in the established learning theory, and monitors weekly engagement trajectories to trigger voluntary player type re-assessment when sustained motivational drift is detected across a three-week cycle. Implemented as a Moodle plugin and evaluated across three university courses involving 135 students, the system successfully differentiated all six Hexad player types and the re-assignment mechanism operated with reasonable selectivity, identifying only students exhibiting persistent engagement decline rather than transient fluctuations. These findings provide proof-of-concept support for treating learner motivation as a dynamic trajectory and suggest that adaptive gamification systems may benefit from combining structured player type differentiation with learner-initiated re-alignment mechanisms.

ABSTRACT. Uncertainty design is considered an important element in promoting engagement and curiosity in digital game-based learning (DGBL). However, prior research has often treated uncertainty as a binary design factor, focusing mainly on whether uncertainty is present or absent, while paying limited attention to how different uncertainty mechanisms are structured. In particular, little is known about how such mechanisms shape different forms of curiosity or how learners’ prior knowledge may condition these effects. To address this gap, this study examined two design-level uncertainty conditions that differed in information transparency and event predictability: predictable uncertainty (PU) and unpredictable uncertainty (UU). An experimental study was conducted with 55 undergraduate students using “Dream Island”, a selfdeveloped game for learning Japanese onomatopoeia. Participants were assigned to either the PU or UU condition, and subgroup comparisons were further conducted based on prior knowledge level. The results showed no significant differences between the PU and UU conditions on overall curiosity measures. However, within the UU condition, learners with low prior knowledge reported higher deprivation-type curiosity than those with high prior knowledge, whereas this pattern was not observed in the PU condition. Interest-type curiosity showed no clear between-group differences. In addition, participants in both conditions showed significant and comparable improvements from pretest to posttest. These findings suggest that the effects of uncertainty design are not uniform, but vary according to curiosity type and learners’ prior knowledge. The study offers evidence for a more mechanism-oriented and personalized approach to DGBL design.

ABSTRACT. Serious games are widely advocated for environmental education, yet few are designed rigorously or measured in ways that preserve engagement. This paper presents Urban Choices, a browser-based narrative game for India's undergraduate environmental studies curriculum, designed around the Design-Play-Experience (DPE) framework to make measurement an integral part of play rather than a post-hoc add-on. The game operationalises UNESCO's eight sustainability competencies through eight scenes spanning a character's first three months in a shared apartment. A cascade architecture embeds early consumption decisions as hidden consequences that surface later, operationalising systems thinking without explicit instruction. The theoretical framework-DPE, ICAP explanation probes, and competency-aligned learning design divides labour among disciplines (game design, learning science, narrative) by giving each a lane while making trade-offs explicit. A pilot playtest with 43 undergraduates found that (1) in-flow telemetry captured every research-critical event with 97.7% completeness and 100% scene-to-capstone retention, demonstrating that unobtrusive measurement and engagement can coexist; (2) players' pre-play framing persisted into later unrelated decisions at well-above-chance rates, supporting the measurement of prior knowledge activation; and (3) explanation depth rose across three ordered probes from active to constructive reasoning, consistent with knowledge construction observable inside gameplay. The paper contributes a design pattern hidden consequences, measurement folded into fiction, discipline-bridged architecture applicable beyond this curriculum, and releases the artifact and telemetry schema openly for re-authoring.

ABSTRACT. Large language models are increasingly explored as interactive tutors for mathematics learning, making it important to understand how model adaptation and Socratic scaffolding each contribute to tutoring quality. This study reports a controlled quality–cost comparison of four Socratic AI math tutor conditions across 241 held-out algebra and geometry problems. The system prompt, student simulator, dialogue protocol, and evaluation framework were held constant. Overall tutoring quality was closely clustered across the four model-adaptation and deployment conditions, and none of the planned contrasts reached practical significance. In contrast, removing the Socratic scaffold produced a substantially larger decline in tutoring quality. Cost analysis likewise showed no clear quality–cost advantage for additional model adaptation or frontier API deployment under the study assumptions. These findings suggest that, in simulated tutoring settings and under the tested prompt design, scaffold design may be a more immediately impactful design lever than model adaptation.

ABSTRACT. Lesson induction, a learning strategy that enables learners to learn from their errors in problem solving, has been shown to be effective. However, deriving high-quality lessons independently remains challenging for many learners. To address this issue, this study proposes and develops an interactive lesson induction support system using LLMs. The system recognizes learners’ handwritten answers and guides their reflection through a four-stage dialogue based on cognitive counseling. A practical evaluation in mathematics with junior high school students showed that the quality of the lessons significantly improved among learners who completed the dialogue through the final stage.

ABSTRACT. Web-based investigative learning is a significant form of self-directed information exploration, in which learners construct knowledge by investigating Web resources. However, learners’ interest often declines as investigation progresses, which also cause an engagement decrease and premature termination of investigation. To address this issue, this paper proposes a method that uses generative AI to present unexpected information tailored to each learner. The proposed method is designed to elicit both cognitive conflict and surprise by presenting information that is difficult for the learner to explain based on what they have learned so far. We have implemented this method in a cognitive tool called iLSB (interactive Learning Scenario Builder), where learning scenarios created by learners and related logs are analyzed to estimate learners’ knowledge learned and to generate unexpected information adaptively. A case study was conducted to evaluate the proposed method. The results suggest that the presented unexpected information can induce surprise and cognitive conflict, enhance learners’ engagement, and promote subsequent Web-based investigative learning. These findings indicate the potential of generative AI not only as a tool for providing answers but also as a means of reactivating information investigation.

ABSTRACT. Open-ended student feedback is an important source of data for improving teaching quality, curriculum design, and learning services. However, these comments are usually written as free-form text, making them difficult to analyze manually at scale. Existing studies on Vietnamese student feedback mainly focus on sentiment analysis or topic classification, while less attention has been paid to identifying issues that can be turned into concrete improvement actions. This paper proposes an action-oriented approach to student feedback analysis with four tasks: sentiment classification, topic classification, actionability detection, and issue type classification. Based on the existing sentiment and topic labels, we construct two additional label groups, namely Actionability and Issue Type, using a rule-based labeling strategy followed by manual review. Using a dataset of more than 16,000 Vietnamese student comments, we compare traditional machine learning models, single-task Transformer models, and multi-task Transformer models. The experimental results show that XLM-R-MTL achieves F1-scores of 0.9413 for sentiment classification, 0.8927 for topic classification, 0.9639 for actionability detection, and 0.9045 for issue type classification. These results suggest that multi-task learning is an effective approach for jointly capturing descriptive and action-oriented signals in student feedback, thereby supporting educational quality assurance in higher education.

ABSTRACT. This study evaluates the pedagogical viability of LLM-generated content within a localised English as a Foreign Language (EFL) context. Utilising log data from Japanese junior high school students practicing on a grammar drilling application, we analysed how different question modalities impact student performance and whether theoretical localised CEFR difficulty tiers accurately predict empirical task difficulty. Results reveal a clear performance hierarchy: multiple-choice questions carried the lowest cognitive load, cloze tasks posed the greatest barrier to active recall, and drag-and-drop exercises incurred the heaviest time penalties. Furthermore, learner data validated the CEFR-J grammar framework, showing a steady decline in accuracy and increased response times as proficiency levels advanced. These findings demonstrate that LLMs can successfully generate learning content, while highlighting the need for developers to strategically sequence question modalities to transition learners from passive recognition to active linguistic production.

ABSTRACT. Artificial intelligence agents are increasingly shaping technology-enhanced language learning by supporting dialogue, feedback, assessment, scaffolding, and personalized practice. However, the field remains conceptually fragmented because related systems are often described using overlapping terms such as chatbots, conversational agents, pedagogical agents, intelligent tutoring systems, dialogue systems, virtual agents, and generative AI tools. This scoping review maps SSCI-indexed Web of Science research on AI agents in language learning to clarify publication trends, agent types, language-skill coverage, learning outcomes, research designs, and future directions. Guided by PRISMA-ScR and JBI-informed principles, the review synthesizes literature on AI-agent-supported language learning across educational technology, applied linguistics, and computer-assisted language learning. The findings show that research has expanded rapidly with the emergence of generative AI and large language models, shifting the field from earlier dialogue-based and tutoring systems toward more flexible conversational, feedback-oriented, and learner-support agents. AI agents are commonly used as speaking partners, writing assistants, feedback providers, personalized tutors, pronunciation-support systems, and self-regulated learning companions. Despite these developments, the field still needs clearer conceptual definitions, stronger longitudinal evidence, deeper attention to ethical AI literacy, more explicit teacher-AI collaboration models, and broader research in multilingual and mobile-first EFL contexts. The review concludes that AI agents should be understood not only as technological tools but also as pedagogical systems that require careful instructional design, learner training, teacher mediation, and responsible integration.

ABSTRACT. Federated learning enables privacy-preserving analysis of educational data across multiple institutions, but its performance is often degraded by non-independent and identically distributed (non-IID) data across clients. Existing studies have primarily addressed this challenge through aggregation strategies, whereas representation-level approaches remain underexplored. This study investigates differential features, which represent pairwise differences in learning-log features and grade-related labels within each client, as a representation-level approach for mitigating non-IID effects. Using learning log data from 17 courses involving more than 1,400 students, we evaluate the impact of differential features on three types of data heterogeneity: label skew, feature skew, and quantity skew. Experimental results show that differential features substantially reduce distributional discrepancies across clients: label JSD decreases from 0.37 to 0.25 and HD from 0.39 to 0.26, while feature JSD decreases from 0.63 to 0.40 and HD from 0.59 to 0.36. Ranking-based at-risk student prediction also improves on held-out clients. Although pairwise sampling amplifies quantity skew through quadratic sample growth, these findings demonstrate the potential of differential features to mitigate label and feature skew in educational federated learning.

ABSTRACT. Concept mapping is a knowledge-building activity that requires learners to interpret concepts, articulate relationships, and integrate new information, yet little is known about how the process of building a map relates to cognitive engagement and outcomes. In this study, we examined how behavioral engagement relates to perceived cognitive load and learning performance, and used transition network analysis to characterize action dynamics. Results show no straightforward linear relationships among behavioral engagement, cognitive load, and performance. Action sequence analysis further revealed that page flipping and object moving dominated students’ behaviors, potentially overshadowing more meaningful action patterns. These findings highlight the need to elaborate the pedagogical framework to model student behaviors and develop support or interventions to encourage productive engagement during concept mapping.

ABSTRACT. This study compared standard error (SE) and confidence interval (CI) stopping rules in licensure computerized adaptive testing (CAT) using a simulation design that varied item difficulty, item discrimination, and examinee ability distribution. The SE rule produced shorter tests when item discrimination was at or above the medium level, whereas the CI rule yielded higher classification accuracy and lower error rates across conditions, despite requiring more items near the cut score. These findings suggest that the SE rule may be useful when test length must be constrained, while the CI rule is more appropriate when pass/fail classification validity is prioritized.

ABSTRACT. Programming assignments are important learning materials in introductory programming education, and computational representations of assignments can support tasks such as curriculum analysis, assignment organization, and similarity-based retrieval. However, programming assignment similarity is not uniquely defined: two assignments may be considered similar with respect to their overall solution structure, loop usage, conditional branching, required data structures, or other aspects. Existing code-based approaches often represent assignments from a fixed perspective, making it difficult to control which structural aspect is emphasized in comparison. This study proposes an AST-based probabilistic representation of programming assignments that enables controllable similarity analysis from selected syntactic perspectives. We evaluate the proposed representation using 47 assignments and learner solutions from an introductory Python programming course. The results show that changing the focal syntactic element produces different and interpretable similarity relationships among assignments. An expert evaluation by the course instructor further suggests that the proposed representation can identify assignment relationships that are meaningful within the target course context. These findings indicate that the proposed representation provides an interpretable and controllable way to analyze programming assignments through solution-code structure.

ABSTRACT. This study examines whether a teachable agent (TA) moderates the influence of mind wandering (MW) tendency on constructive and interactive (C/I) activities in computer-supported collaborative learning (CSCL). A total of 131 university students participated in pairs and were randomly assigned to either a TA or no-TA condition using CoCot, a collaborative concept mapping system. In addition to the TA manipulation, metacognitive suggestions were provided under all conditions as a baseline form of collaborative process support. MW tendencies were assessed using the Daydream Frequency Scale (DDFS), and chat logs from the collaborative phase were coded according to the ICAP framework (Chi & Wylie, 2014). The results indicated that the TA condition produced significantly higher proportions of constructive and interactive utterances and a lower proportion of active utterances than the no-TA condition. Regarding the relationship between MW tendencies and C/I activity, no significant correlations were found in the TA condition, whereas positive correlations between DDFS scores and C/I utterance proportions emerged in the no-TA condition. These findings suggest that a TA exerts an equalization effect on C/I activity—homogenizing engagement, regardless of individual MW tendencies, while also revealing that in the absence of a TA, learners with higher MW tendencies exhibit greater spontaneous generative and interactive engagement. The implications of these findings for CSCL system design and for supporting learners with diverse MW profiles are discussed.

ABSTRACT. Collaborative learning remains valuable in the AI era as the social-interactional aspects of learning are still irreplaceable, while its classroom implementation remains challenging due to the burden group activities can place on teachers. Though data-driven technologies have increasingly supported individual learners through explainable intelligent systems, interpretable and actionable AI support for group learning remains underexplored, especially in K12 classrooms. To bridge this gap, this study proposes an explainable data-driven group learning paradigm where LLM-based narratives complement algorithmic group formation and group awareness via Wordcloud. A pilot study was conducted with second-year Japanese senior high school students from basic and advanced EFL classes using a crossover design. The results showed an inconsistent objective advantage of the narratives and might suggest additional cognitive load for higher-performing learners. Lower-level learners perceived the support as more helpful, indicating the feasibility and conditional value of translating learner analytics into human-readable support for classroom group learning.

ABSTRACT. Makerspaces are failure-rich environments where persistence is central to how learners continue engaging with open-ended projects. Research on persistence in maker education documents that students spend substantial time on projects, return across sessions, and describe their experiences as ones of overcoming difficulty. However, existing approaches treat challenges or obstacles learners face as equivalent across instances. In collaborative electronic making, this assumption does not hold. Challenges vary fundamentally and make qualitatively different persistence demands. To address this gap, we conducted a three-week workshop on collaborative electronic making (N=20). A total of 152 trouble episodes were identified and analysed from approximately 51 hours of video documentation using interaction analysis. Episodes were coded inductively for trouble type, and persistence was characterised through attempt count and episode duration. Inductive coding produced six trouble types and revealed that digital and hardware troubles dominated the trouble landscape. The number of attempts per episode established a distinction between bounded troubles and open troubles. Plotting episodes against attempt count and mean duration per attempt revealed qualitatively different persistence profiles. A wheel-spinning-like profile appeared almost exclusively in open troubles and was associated with a 60% unresolved rate. A deliberate, sustained engagement (productive persistence) profile was associated with a substantially lower unresolved rate (33.3%). These findings suggest that trouble type is a necessary context for interpreting persistence metrics in open-ended electronic making. The trouble typology developed here offers researchers a systematic basis for studying persistence and provides a conceptual framework to inform facilitator intervention decisions.

ABSTRACT. Intelligent tutoring systems are routinely tested before classroom deployment with simulated students, yet most simulators model only cooperative learners. This leaves a systematic evaluation gap, because real tutoring logs also contain students who guess rapidly, stall on a single skill, or disengage for long periods. We present Disengagement-Aware Student Simulators (DAS²), a reproducible pre-deployment framework for auditing whether LLM-based tutors stay trustworthy across the full range of learner engagement. DAS² generates simulated students in five learner states (engaged, gaming, wheel-spinning, off-task, and mixed) and tests whether tutor scores and rankings are sensitive to those states. In a 100-session blinded audit of ASSISTments09, two independent coders reached substantial agreement (κ = 0.78; 84% agreement), and human-agreed labels aligned with rule-assigned labels (κ = 0.75), supporting their use as evaluation groups. Conditioning simulation on these labels reduced the simulated-to-real gamer correctness gap from 0.54 to 0.20, and a post-hoc calibration step narrowed the gaming and wheel-spinning correctness gaps to at most 0.05. Across 10-turn and 20-turn Arena evaluations of five LLM tutors, rankings stayed stable under the primary judge (Spearman ρ = 1.00), yet absolute judge scores varied meaningfully across learner states, and a weak-tutor injection recovered the expected order, placing GPT-4o last on engagement with non-overlapping confidence intervals. A transparent construct-validity audit lifted the weak relevance judge from ρ = 0.33 to ρ = 0.61, though composite agreement reached ρ = 0.82, below the pre-registered 0.85 target. We treat this score-versus-rank dissociation as the central finding: a stable leaderboard can mask qualitatively different levels of tutor support for disengaged students. DAS² offers an auditable, reproducible way to inspect learner-state coverage before deployment on a single dataset and judge; it does not measure classroom learning gains.

ABSTRACT. Recently, as Large Language Models (LLMs) have fundamentally impacted various fields, the methods for incorporating up-to-date information into LLMs or adding external knowledge to construct domain-specific models have garnered wide attention. Retrieval-Augmented Generation (RAG), serving as an inference-time scaling method, is notable for its low cost and minimal effort for parameter tuning. However, due to heterogeneous training data and model architecture, the variant embedding models used in RAG exhibit different benefits across various areas, often leading to different similarity calculation results and, consequently, varying response quality from LLMs. To address this problem, we propose and examine two novel approaches that combine the benefits of multiple embedding models, named Mixture-Embedding RAG and Confident RAG. Mixture-Embedding RAG simply sorts and selects retrievals from multiple embedding models based on standardized similarity; however, it does not outperform vanilla RAG. In contrast, Confident RAG generates responses multiple times using different embedding models and then selects the responses with the highest confidence level, demonstrating average improvements of approximately 10% and 5% over vanilla LLMs and RAG, respectively. The consistent results across different LLMs and embedding models indicate that Confident RAG is an efficient plug-and-play solution for mathematics question answering. The code of our paper is provided at https://github.com/RS2002/Confident-RAG.

ABSTRACT. This study examines the dual role of generative AI (GAI) in a cross-disciplinary higher education course. On the instructor side, GAI was investigated as a calibration reference within a tri-rater framework involving Teacher A, Teacher B, and an AI Rater. On the student side, it was examined as a learning scaffold within a human–AI collaboration context. Adopting a repeated cross-sectional field study design, the broader study spans three consecutive semesters and is expected to include 297 students in total; however, the current submission reports analyses based on the two completed tracks available at the time of submission. Bland–Altman analysis and regression models were used to examine rating consistency and students’ technology acceptance. The findings indicate significant scoring discrepancies between the two human teachers, with both the direction and magnitude of these discrepancies varying across semesters. This pattern suggests the presence of systematic bias and rater drift, highlighting the difficulty of ensuring scoring consistency through one or two human raters alone. In the tri-rater analysis, the AI Rater did not function as a fully neutral scorer; however, its scores often provided a useful reference point for examining disagreement among raters, particularly when divergence between the two teachers was more pronounced. At the same time, the AI Rater showed larger discrepancies in some humanities-related tasks, suggesting that its performance remained sensitive to task characteristics. Overall, the magnitude of AI-related discrepancies did not appear to exceed the range of disagreement observed between human raters, supporting its potential value as a supplementary assessment tool. Further analysis showed that perceived ease of use predicted perceived usefulness, whereas system satisfaction depended primarily on perceived usefulness rather than ease of use alone. These findings suggest that the educational value of AI lies not in replacing teachers, but in supporting human–AI collaborative assessment as a practical calibration reference while also functioning as a learning scaffold for students.

ABSTRACT. This study investigates an instructional approach integrating GenAI, embodied cognition (EC), and argumentation-based instruction (ABI) to support junior high school students’ AI learning. Thirty-eight students participated in a two-day program involving conceptual instruction, embodied activities, GenAI tasks, and the AI2 Robot City board game. Results showed significant improvements in students’ learning effectiveness and metacognitive awareness. Behavioral sequence analysis further revealed connections between embodied actions and argumentative discourse, suggesting that students engaged in reasoning, coordination, questioning, and collaborative knowledge construction. As an exploratory study, the findings provide initial evidence for integrating EC and ABI in GenAI-supported learning environments.

ABSTRACT. The widespread integration of AI literacy in higher education remains challenging, particularly when it comes to helping beginners grasp abstract AI related-concepts. Although game-based learning (GBL) has been explored in prior research, its application in compulsory, asynchronous AI courses for diverse undergraduate populations has received limited attention. This study examines a GBL-based AI literacy course by analyzing pre- and post-survey data from 311 first-year undergraduates, focusing on cognitive, behavioral, affective, and ethical learning outcomes. The results indicate significant improvements in cognitive, behavioral, and affective dimensions, but no statistically significant change in ethical learning. By providing large-scale empirical evidence, this study contributes to the literature and proposes instructional design for integrating AI literacy into higher education contexts.

ABSTRACT. This study compared primary school students' flow experiences between 2D game development and its 3D codebase extension. Twenty-nine fourth graders built a 2D maze game in MakeCode Arcade and, one week later, extended the same code into a 3D maze game through a Raycasting extension. Flow was measured after each condition using a nine-dimensional Learning Flow Scale. Overall flow scores were significantly higher in the 3D condition than in the 2D condition, particularly in three of the nine sub-dimensions. The findings suggest that scaffolded 2D-to-3D transitions may offer an effective pathway for sustaining motivation and enhancing engagement in game-based programming education.

ABSTRACT. Micro-course learning offers flexible and personalized learning opportunities, yet learners often experience divided attention and limited engagement. This study examines how Student Activity Sheets (SAS) support attention and learning in micro-course environments through eye-tracking-based learning analytics. Seventy-seven elementary school students were randomly assigned to a worksheet group, a graphic organizer group, or a control group without SAS. Eye-tracking indicators, including first fixation time, number of long fixations, and total fixation duration, were used to capture learners’ attentional processes, while retention and transfer tests assessed learning outcomes. Results showed that both SAS formats significantly improved learners’ attention and learning performance compared with the control condition, with no significant differences between the two SAS formats. Mediation analysis further revealed that attention, particularly total fixation duration, mediated the relationship between SAS use and learning outcomes. These findings provide process-level evidence for the role of SAS in scaffolded micro-course learning and demonstrate the value of integrating eye-tracking analytics with learning assessment to understand attentional mechanisms in technology-enhanced learning environments.

ABSTRACT. Generative AI (GAI) agents are increasingly used in educational tasks, but less is known about how they redistribute students’ reasoning during learning. This qualitative multiple-case study analyzes interaction logs from two pedagogical agents used in computational-thinking lesson-design tasks. We define AI-mediated thinking as the interactionally visible ways learners’ reasoning, judgment, regulation, and agency are supported, delegated, or displaced during student–AI interaction. Using an AI-adapted Activity–Community of Inquiry (A–CoI) framework, episodes were analyzed through activity-theory components and the distribution of cognitive, teaching/guiding, and social presence across students, agents, and generated artifacts. Four modes were identified: delegated completion thinking, co-constructive object formation, regulated object grounding, and elicited reasoning synthesis. The Pedagogical Guided Agent tended to support artifact completion or co-construction, depending on student contribution, whereas the Socratic Questioning Agent delayed production, gated progress, and elicited students’ scenario details, variables, criteria, and decision rules before synthesis. Findings show that the educational meaning of GAI use depends on how agents configure the object of activity, participation rules, division of labor, and presences. Educational GAI agents should sequence generation after problem grounding, student-first elicitation, and pressure testing so that AI output synthesizes rather than replaces student reasoning.

ABSTRACT. In face-to-face communication, individuals infer others' emotions not only from verbal content but also from nonverbal cues such as facial expressions, voice, and gestures. However, existing training methods for emotion recognition focus on single, clearly expressed emotions, despite real-world interactions often being complex and partially concealed. In this paper, we define emotional expressions in which positive emotions conceal negative ones as Compound Emotional Expressions. We also design robot-based emotional expressions to represent simple and compound emotions in a human-like manner, in which a social robot combines nonverbal behaviors including facial expressions, voice, and gestures to generate both single and compound emotional expressions. We conducted a case study to evaluate whether designed single and compound emotional expressions were recognized as intended. The results showed high recognition accuracy for single emotional expressions. Compound emotional expressions were also recognized to some extent and perceived contextually plausible. These findings indicate the possibility of representing compound emotions with a social robot through combined multiple nonverbal behaviors.

ABSTRACT. This study constituted a qualitative multiple-case analysis of how Artificial Intelligence (AI) is integrated across two educational contexts: institutional-pedagogical implementation in secondary schools and cognitive-mediated learning in higher education. Grounded in a multiple-case design, the study combined within-case and cross-case analysis to explore how AI functions across these different educational environments. The first case study examined 12 Arab teachers and 12 Arab principals’ perceptions of Artificial Intelligence in Education (AIED) integration into Arab secondary schools in Israel. The second case study explored the experiences of 27 graduate students using AI-based pedagogical chatbots and scenario-based simulators in self-directed learning environments. Findings show AI was perceived differently across contexts: in secondary schools, AI emerged as an organizational and pedagogical innovation requiring leadership, infrastructure, and ethical regulation, whereas in higher education, it functioned as a cognitive mediator supporting reflection, scaffolding, self-regulation, and autonomous learning. The findings suggest AIED represents a multidimensional educational transformation operating simultaneously at organizational, pedagogical, and cognitive levels.

ABSTRACT. Instructional coaching can support teachers to notice, interpret, and improve classroom practice, but access to expert coaching remains uneven because coaching is time-intensive and costly to scale. Generative artificial intelligence (GenAI) may offer a complementary mechanism for providing timely, rubric-aligned feedback. This exploratory study examines whether a GenAI-based instructional coach can approximate selected aspects of expert human coaching judgment when all coaches analyze the same classroom evidence under the same pedagogical framework. Two expert human coaches and one GenAI coach independently reviewed audio recordings from three classroom lessons in a K-12 school in Mumbai, India. All feedback was structured using Danielson's Framework for Teaching, Domain 3: Instruction. We compared feedback volume, issue-level convergence, scoring patterns, uniqueness of feedback, and time and cost requirements. Results show that the AI coach generated a comparable number of feedback points to one human coach, that 65% of clustered instructional issues were common across human and AI feedback, and that AI average scores were within the human-score range in two of three lessons. In the third lesson, the AI score was slightly above the human range. AI-generated feedback was substantially faster and lower cost than human coaching. These findings do not establish coaching effectiveness or substitutability. Rather, they suggest that, under constrained audio-only and rubric-aligned conditions, GenAI can participate in instructional noticing and first-pass feedback generation while requiring human oversight for contextual interpretation, prioritization, ethical use, and teacher uptake.

ABSTRACT. This paper investigates the effectiveness of cluster classroom design on non-cognitive learning outcomes among foundation-level students in a Malaysian higher education context. Using a quasi-experimental pretest–posttest design, the study examines three key dimensions of a cluster classroom, namely, physical environment, teacher support, and social climate. Results from paired samples t-tests indicate significant improvements in teacher support and social climate within the experimental group. However, a one-way ANCOVA revealed no statistically significant difference in overall non-cognitive outcomes between the cluster and traditional classrooms. These findings indicate that changing physical layout alone does not cause significant improvement in students’ non-cognitive outcomes. Instead, the value of a cluster classroom lies in its role as a structural catalyst that successfully enables more responsive teacher support and collaborative social scaffolding.

ABSTRACT. This study investigates the impact of AI-supported formative feedback on student report revision in higher education. Although prior research has primarily focused on text-based assignments, reports used in authentic classroom settings often include multimodal elements such as slides, figures, and visual explanations. The effectiveness of AI-supported feedback for such multimodal reports remains underexplored. To address this gap, we developed an LLM-based system that analyzes multimodal reports through a unified prompt design and generates formative feedback to support students’ self-reflection before submission. The system integrates structure detection, content summarization, and suggestion generation within a single multimodal large language model. We conducted a comparative study in a graduate-level course with more than 100 students, comparing two academic years: one without AI support (2024) and one with the AI system (2025). Multiple data sources were analyzed, including peer evaluations, instructor evaluations, report artifacts, system usage logs, and student surveys. The results show that report quality was significantly higher after the introduction of AI-supported feedback (p < .001), with particularly strong gains in higher-order components such as discussion and conclusion. Students also increased the amount of content in their reports, suggesting more constructive revision behavior. Survey results further indicate generally positive perceptions of the system. These findings provide empirical evidence that LLM-based formative feedback can support the revision of multimodal reports in authentic classroom settings and highlight its potential as a scalable support tool in higher education.

ABSTRACT. Generative AI (GenAI) is reshaping project-based and learning-by-doing courses by offering rapid procedural assistance, yet unrestricted support may encourage answer offloading and weaken teachers’ ability to orchestrate synchronous classroom learning. This study proposes a Hierarchical Digital Twin Ecology, a teacher-governed multi-agent architecture in which student-facing digital mentors provide step-based guidance, request evidence before escalating assistance, and aggregate real-time learning states to a teacher dashboard for timely classroom intervention. Its key mechanism is a state-aggregation and threshold coordination protocol that turns stalled progress, missing evidence, repeated failures, and help-seeking patterns into visible indicators for teacher noticing and synchronous intervention. Rather than advancing a stand-alone psychological theory, the study operationalizes identity-based motivation, the zone of proximal development, and the Mantle of the Expert as proximal identity scaffolding for evidence-first recovery. A 14-week quasi-experimental study compared three conditions in a robotics and Artificial Intelligence of Things (AIoT) project course with 120 students: LMS-only support, Butler-style execution support, and the proposed Digital Twin Ecology. Because intact classes were used, baseline project performance was examined using one-way ANOVA, and outcome comparisons used ANCOVA with prior project performance as the primary covariate. For RQ1, the ecology group showed the strongest evidence-first recovery profile, including the highest 48-hour recovery after setbacks (72.86%), evidence checkpoint completion (85.61%), evidence submission before help requests (69.41%), reattempts after setbacks, and debugging-note completion. For RQ2, the ecology group achieved the highest project performance (M = 87.45), especially in debugging quality and documentation/evidence explanation, indicating stronger diagnostic and evidential competence. For RQ3, qualitative records and student reflections suggested that recovery was supported by evidence-first help seeking, proximal role framing, and dashboard-triggered teacher intervention when shared bottlenecks became visible. By linking evidence checkpoints with teacher-facing analytics, the ecology translated individual recovery traces into actionable classroom coordination. Together, these findings suggest that teacher-governed GenAI can support individualized assistance while preserving students’ epistemic agency and strengthening classroom-level orchestration. The study contributes a practical design account of teacher–AI complementarity in project learning and offers a transferable coordination logic for AI-supported project classrooms, while acknowledging the limits of its single-course, intact-class design and context-specific 48-hour recovery indicator.

ABSTRACT. Project-based learning (PBL) is widely recognized for fostering higher-order thinking in programming education, yet its effectiveness is often constrained by insufficient scaffolding and feedback. With the emergence of generative artificial intelligence (GenAI), new opportunities arise to support learners during complex project tasks. This study proposes a Generative AI-supported Project-Based Programming Learning (Gen-PPBL) model and investigates its effectiveness through a quasi-experimental study involving 81 high school students. The experimental group received GenAI-supported instruction, while the control group followed conventional PBL. The results showed that the experimental group significantly outperformed the control group in decision-making, creative thinking, critical thinking, and problem-solving abilities. Students also demonstrated improvements in project innovation, learning engagement, and programming confidence. The findings suggest that GenAI functions as an effective cognitive scaffold, enhancing student engagement and supporting complex problem-solving in programming education.

ABSTRACT. Generative AI learning companions are increasingly used in programming education, but students may still struggle to transform AI feedback into conceptual understanding and confidence. This study presents TaskInsighter-CM, a generative AI-supported learning companion that combines conversational self-explanation scaffolding with concept map–embedded error diagnosis. The system prompts students to explain their reasoning, identify conceptual errors in Python-related concept maps, receive rubric-based feedback, and revise their explanations. A quasi-experimental pretest–posttest study was conducted in a graduate Python course in Taiwan with 52 students. The experimental group (N = 35) used TaskInsighter-CM, while the control group (N = 17) followed the regular instructional arrangement. After controlling for pretest scores, the experimental group showed higher Python achievement and overall programming self-efficacy than the control group, with significant differences in Cooperation, Control, and Debug self-efficacy. Students also reported positive perceived usefulness and ease of use. The findings provide preliminary evidence that combining generative AI dialogue with concept map–based error diagnosis may support programming learning and process-oriented programming self-efficacy.

ABSTRACT. Although self-regulated learning (SRL) is generally assumed to promote persistence and achievement, highly self-regulated learners do not always maintain stable engagement in online learning environments. This study investigated the temporal relationship between SRL, perceived difficulty, and engagement collapse using longitudinal weekly learning records from university EFL learners (417 learners, 2,970 learner-week observations). Linear mixed-effects modeling revealed a significant negative relationship between SRL and completion rates. This relationship was initially moderated by perceived difficulty, such that higher SRL predicted lower completion more strongly under lower perceived difficulty conditions. However, robustness analyses excluding complete non-participation episodes (0% completion) showed that the interaction disappeared, suggesting that the observed effect primarily reflected emerging disengagement rather than active learning behavior itself. To examine temporal disengagement processes, survival analyses were conducted using individualized baseline engagement levels. Cox proportional hazards modeling indicated that higher SRL significantly predicted earlier collapse timing (HR = 1.39, p < .001), while learner clusters demonstrated distinct collapse risks and engagement trajectories. These findings suggest that SRL-related reflective activity may function as a context-sensitive and temporally reactive signal associated with engagement instability and impending disengagement in online learning environments.

ABSTRACT. This study proposes a cognitive-model-based workflow for generating AI-based visual resources to support literary reading. Existing educational applications of generative AI have often emphasized practical efficiency rather than the cognitive processes involved in literary comprehension. To address this issue, the study introduces “visual resources” as visual information supporting readers’ construction of meaning during reading. The proposed workflow integrates situation model theory, Dual Coding Theory, and human review to connect literary analysis, drawing policy generation, and image generation. The workflow emphasizes balancing cognitive support and interpretive openness so that generated images assist emotional inference without excessively constraining interpretation. Through an exploratory application to Japanese literary texts and a lexical analysis of teachers’ impressions, the study discusses the educational possibilities and limitations of AI-generated visual resources in literary education.

ABSTRACT. Large Language Model (LLM)-driven Socratic dialogue may deepen cognitive engagement in Robot-Assisted Language Learning (RALL), yet its usability cost across distinct facets remains underexplored. This exploratory study deployed three AI agent designs—Direct Feedback, Socratic, and Adaptive Socratic—on a Robot-and-Tablet (R&T) platform with 32 ninth-grade English as a Foreign Language (EFL) students in Taiwan. A multidimensional usability analysis revealed significant group differences across all three dimensions—Usefulness, Easiness, and Satisfaction—with the Socratic agent consistently scoring the lowest; the deficit was most consistent on Usefulness, the only dimension on which both pairwise contrasts remained significant following a Bonferroni correction. Decomposing Easiness revealed an asymmetric pattern: the deficit was largely confined to Ease of Use (p = .012), with no significant difference observed in Ease of Learning (p = .163); because the platform interface was held constant across conditions, this asymmetry locates the cost in the Socratic interaction pattern rather than the shared interface. The Adaptive Socratic agent (which directed challenges to unmastered items via cumulative error-tracking) did not differ significantly from the Direct Feedback agent across any usability dimensions; however, this null pattern cannot establish equivalence at the present sample size, and direct contrasts between the Socratic and Adaptive Socratic agents were only marginally significant after statistical correction. Adaptive targeting is therefore advanced as a design hypothesis for mitigation rather than as an established empirical effect. Overall, students' English learning attitudes improved significantly (p = .001), with no agent-specific differences detected. Together, these findings reframe Socratic difficulty as a dimension-specific usability cost in AI-tutored RALL, selectively reducing Ease of Use with no detectable effect on Ease of Learning, and motivate further investigation of adaptive content targeting as a candidate mitigation.

ABSTRACT. Universities and online learning platforms generate student records used for research on dropout prediction, knowledge tracing, and adaptive tutoring. Direct sharing of these records is restricted by privacy regulations, including FERPA (Family Educational Rights and Privacy Act) and GDPR (General Data Protection Regulation). Differentially private (DP) synthetic data, which replaces real records with generated substitutes that bound each individual's influence on the release, offers one path to compliant sharing. However, existing evaluations report privacy risk and utility as separate scores, missing the practical release question: a dataset may pass a privacy check while losing the information needed for the intended task. We introduce PEARL (Privacy-Equivalence Audit and Release Ledger), a fail-closed audit benchmark for DP synthetic educational data. PEARL rejects a candidate release unless it passes sequential checks for validity, privacy, utility, and task-specific educational use, and records the reason each rejected release fails. Applied to 96 configurations across tabular student records, MOOC clickstreams, and knowledge-tracing datasets, PEARL certifies 12 releases. These include nine tabular releases from MST or DP-CTGAN and three ACT-MOOC releases from DP-Markov (membership inference attack AUC = 0.491, retention = 0.922 at ε = 1.0). Many privacy-passing releases fail due to missing target classes, low utility, label artifacts, or loss of ordered skill-transition structure. A subgroup fairness gate additionally rejects tabular releases that distort at-risk recall across protected groups, and stronger shadow-model and likelihood-ratio attacks leave every certification unchanged. Deep Knowledge Tracing (DKT) and Self-Attentive Knowledge Tracing (SAKT) achieve exactly 0.000 F1 on synthetic data across all tested settings (five seeds, every privacy budget, every synthesizer). This is a reproducible sign that sequence structure is destroyed rather than an implementation artifact. PEARL shows that synthetic educational data should be judged by whether each release still supports its intended educational use, such as knowledge tracing, dropout prediction, and adaptive tutoring, not by privacy alone. Audited this way, responsible data sharing and learner data rights advance, rather than undercut, research in intelligent learning environments.

ABSTRACT. Conventional Learning Analytics (LA) has delegated the design of assessment criteria to institutions and researchers and has relied on Learning Management System (LMS) logs, thereby relegating learners to passive consumers of analytical results and excluding practical experience data—such as internships and project-based activities—from analysis. This study proposes an LA system that leverages Large Language Models (LLMs) to enable learners to design assessment criteria, jointly analyze formal learning data and practical experience data, and derive design principles from an evaluation study (N=12) using a prototype implementation. The findings indicate that (1) institutionally collected course data serve as an initial reference frame for self-understanding, but assessing interpersonal, self-management, and problem-solving skills requires the addition of practical experience data; and (2) generic skill axes and custom skill axes complement one another—the former supporting retrospective reflection of past experiences and the latter supporting future-oriented reflection. This work contributes design principles for designing LLM-supported, learner-centered LA systems in which learners participate in constructing both the data and the assessment axes used for reflection.

ABSTRACT. Knowledge monitoring (KM), the ability to judge what one knows and does not know, is a core metacognitive process in self-regulated learning. Although the Knowledge Monitoring Assessment (KMA) framework provides a theoretically grounded approach for assessing KM, limited work has examined how to translate KMA-based information into learner-facing tools for metacognitive support. This study presents the design and formative evaluation of MetaCompass, an AI-based dashboard grounded in KMA framework. MetaCompass is designed through three core principles: dual information visualization of performance and monitoring accuracy, interpretable learner profiles, and profile-aligned adaptive feedback. A formative evaluation was conducted with undergraduate students and instructors to examine the validity of the feedback design, usability, and learning usefulness of MetaCompass. The results indicated that while MetaCompass was perceived as usable, perceived learning usefulness varied among students. The correlation analysis showed that usability was positively correlated with perceived learning usefulness only among students with high metacognitive awareness. These findings indicated that usability alone is insufficient to ensure learning value and highlight the need for metacognitive scaffolding to help learners interpret dashboard data, providing design-based insights into the conditions under which learner-facing dashboards can support KM.

ABSTRACT. This study presents the development and evaluation of a personality-based, serendipity-oriented location recommendation chatbot. While serendipity has been widely discussed in recommender system research, few systems have operationalized personality traits as a mechanism to intentionally induce serendipitous experiences. To address this gap, we designed a chatbot that incorporates users’ Big Five personality traits (TIPI-J) into the recommendation process. Specifically, one personality dimension was inverted to generate recommendations that partially deviate from users’ typical preferences while maintaining contextual relevance. The system was implemented using a large language model (LLM) to dynamically generate recommendations based on personality traits and users’ prior location selections. An experimental study was conducted with 38 university students, comparing a serendipity-oriented condition (n = 20) with a non-serendipity control condition (n = 18). Participants evaluated recommended locations in terms of interest, unexpectedness, perceived serendipity, and willingness to visit. Results showed that although differences in interest and perceived serendipity did not reach statistical significance, medium effect sizes were observed for unexpectedness and serendipity. Importantly, willingness to visit was significantly higher in the serendipity-oriented condition (Mann–Whitney U = 116, p = .016), with a moderate effect size. Ordinal logistic regression further indicated that group membership significantly predicted visit willingness even after controlling for personality traits, suggesting that the observed effect was attributable to the system design rather than individual personality differences. These findings demonstrate the feasibility of inducing serendipitous experiences through personality-based design strategies in chatbot-driven recommendation systems. The study contributes to the development of LLM-based recommender systems that move beyond preference matching toward intentional experience design.

ABSTRACT. This paper examines the deployment of Team-Based Learning (TBL) in teaching Artificial Intelligence (AI) through two case studies in large university classes: an interdisciplinary undergraduate course and a technical postgraduate engineering course. Utilizing TBL as the core pedagogical framework, both courses leveraged readiness assessments, team-based missions, and a purpose-built collaborative environment. By analyzing scores and feedback from students and instructors, we evaluate the efficacy of TBL in these contrasting contexts. Our findings demonstrate that TBL scales effectively for large cohorts while fostering interdisciplinary dialogue, integrating technical and societal perspectives, and cultivating essential soft skills. This work contributes to the field of AI education by illustrating how TBL transcends traditional lecture-based delivery to promote collaborative, ethical, and practice-oriented learning.

ABSTRACT. As generative artificial intelligence (AI) becomes embedded in everyday teaching, pre-service teachers must be prepared not only to use such tools but to integrate them pedagogically. This study examined the effects of a course in which pre-service teachers developed their own instructional materials using generative AI (a "vibe coding" approach) on their Technological Pedagogical Content Knowledge (TPACK) and AI teaching efficacy. A single-group pre-post design was used; of the participants, 20 who completed both the pre- and post-tests were matched and analysed using paired-samples t-tests, Wilcoxon signed-rank tests, and Cohen's d. AI teaching efficacy improved significantly overall (t(19) = −8.83, p < .001, d = 1.97), with the largest gains in AI instructional-design competence (PATE, d = 1.83). TPACK also rose significantly across all sub-domains (overall d = 1.74), with the largest effects in the technology-integrated domains (TPACK, d = 1.57; TPK, d = 1.26; TCK, d = 0.87). General attitudes toward AI (ATSE) and interaction with AI (IWAI) also improved significantly, but with notably smaller effects (both d = 0.55). Findings suggest that authentic AI-based material-development tasks strongly strengthen pre-service teachers' capacity to integrate AI into teaching, whereas broader dispositions toward AI shift to a more modest degree within a single-semester intervention.

ABSTRACT. This paper presents Vesper, an automatic evaluation system for visual elements in statistical posters using convolutional neural networks (CNNs). In high school Informatics education in Japan, the creation of statistical posters is sometimes included in the curriculum. Such posters are evaluated from both data utilization and information design perspectives; however, this system focuses specifically on the visual elements of information design. We construct CNN-based models to evaluate eight visual elements and develop a web-based system that assesses student-uploaded posters and provides feedback with confidence scores. A classroom-based field experiment showed that 97.6% of students rated the system as useful, indicating its effectiveness as a learning support tool.

ABSTRACT. This study examines how an anthropomorphic virtual pedagogical agent (VPA) reorganizes students’ collaborative problem solving (CPS) processes through affective-cognitive support. Rather than treating pedagogical agents only as task tutors, we conceptualized the VPA as a CSCL support mechanism that couples cognitive regulation with socio-emotional cues. A quasi-experimental study was conducted with 126 undergraduate students assigned to three conditions: an anthropomorphic VPA integrating cognitive and affective regulation, a cognitive-only pedagogical agent (PA), and a no-agent condition (NA). The intervention lasted eight weeks and involved collaborative learning tasks aligned with the PISA 2015 CPS framework. CPS outcomes were measured using a 24-item questionnaire, and group discourse was analyzed through metacognitive coding, CPS coding, epistemic network analysis (ENA), and lag sequential analysis (LSA). Results showed that both agent-supported conditions improved overall CPS, but the VPA produced stronger gains than the PA, especially in representation and monitoring. Process analyses further indicated that the VPA strengthened connections among task understanding, strategy selection, planning, monitoring, and emotion/conflict handling, and supported more integrated cross-dimensional CPS sequences. The findings suggest that anthropomorphic features can function as pedagogically meaningful affective-cognitive regulation mechanisms when tightly coupled with task guidance, role coordination, and process monitoring. This study contributes to AI-supported CSCL by explaining not only whether an agent improves CPS, but how anthropomorphic affective-cognitive support reshapes collaborative regulation.

ABSTRACT. Generative AI (GenAI) can help students rapidly create persuasive storyboards, but these outputs may also allow learners to self-confirm imagined designs without testing their fit with physical, contextual, and client needs. This study proposes Metaboard, a generative-AI-supported cyber-physical learning design that transforms storyboards into situated validation spaces. In Metaboard, students enter LLM-generated professional scenarios, define product-context interfaces, synchronize physical prototypes with digital twins, revise prompts when cyber-physical mismatches occur, and publicly demonstrate their designs as future professionals. We conducted a quasi-experimental study in an artificial intelligence of things (AIOT) and robot project course and implemented the same Metaboard principles in a hospitality design course to examine transferability across domains. Results suggest that Metaboard improved learning motivation, design thinking, situated cognition, psychological ownership, and individual project-based learning (PBL) performance compared with GenAI-supported scenario demonstration. Qualitative evidence further suggests that Metaboard shifted students from self-demonstrable functionality toward client-oriented, interface-connected, and publicly accountable design reasoning. The hospitality course further suggested transferability to service design. This study contributes a design principle for AI-supported PBL: AI-generated scenarios become educationally meaningful when they are made answerable to physical constraints, contextual interfaces, professional identity, and public validation through epistemic friction.

ABSTRACT. While peer learning is a highly effective collaborative educational approach, digital platforms often struggle with low participation and unresolved questions. To address these challenges, this paper presents a novel approach within the field of Artificial Intelligence in Education (AIED) by integrating Large Language Model (LLM)-based simulated peers into an online peer learning system. Unlike traditional pedagogical agents that act as authoritative instructors, these data-driven simulated peers function as reciprocal learning companions to provide adaptive learning support within a human-AI co-existing community. Grounded in Open Knowledge and Learner Model (OKLM) which tracks individual learner proficiency from learning logs, the system continuously monitors the community and adaptively intervenes. The simulated peer employs two primary mechanisms: (1) it offers heuristic hints to guide learners through unresolved questions without revealing the final answers and (2) generates tailored simulated help-seeking messages to scaffold passive learners into active teaching roles. Furthermore, an empirical study was carried out in a Japanese junior high school to evaluate the system's impact. By embedding these adaptive AI-driven agents within an existing platform, this framework successfully sustains interaction continuity and collaborative knowledge construction as a supplement of human peers.

ABSTRACT. In organizational knowledge creation, it is important to share and accumulate the insights and judgments that members gain through discussions and to develop them into reusable knowledge within the organization. However, in discussions related to research activities, although outcomes and conclusions are often shared, the tacit understandings underlying them, such as assumptions and criteria for judgment, are not always made explicit. In this study, we developed a support system that records utterances in which tacit knowledge is perceived to have been activated during discussions, formalizes them as Knowledge Fragments after the discussion, and further supports their construction and accumulation as organizational knowledge. An initial practice in a research laboratory suggested that the system has the potential to support activities in which individual insights are developed into organizational knowledge.

ABSTRACT. Coding of learner behavior is a basic methodology in educational research, but its automation in physical-space settings still relies heavily on manual annotation, which constitutes a major bottleneck for scaling up video-based analysis. Classical computer vision and machine learning approaches typically require per-study dataset construction and fine-tuning, which is difficult given that educational data is not always captured in the same environment and due to limited model flexibility. In this study, we propose a method for automatically estimating the object on which a learner is fixating. The method takes as input a first-person video recorded by a wearable eye tracker, with the learner's gaze position overlaid on each frame, and uses a Vision-Language Model (VLM) guided by natural-language class definitions to classify the fixated object. Overlaying the gaze position bypasses gaze-direction inference, at which VLMs underperform, while leveraging their strength in open-vocabulary recognition. We applied the method to a Russian-as-a-second-language reading and writing task performed by five university learners, classifying the fixated object at 10-second intervals into five researcher-defined classes. Compared with human-annotated ground truth, the VLM achieved an accuracy of 0.870 and a Cohen's kappa of 0.781. The results suggest that the proposed method can classify the objects indicated by overlaid gaze markers with substantial agreement with human annotation.

ABSTRACT. This exploratory study re-analyzed student-chatbot programming conversations to examine whether observable interaction patterns differed across learning-gain groups. The dataset comprised 50 students, 1,550 message turns, and 96 conversation sessions. Students were classified according to the direction of change from pre-test to post-test scores: increase (post-test score higher than pre-test; n = 27), same (no score change; n = 14), and decrease (post-test score lower than pre-test; n = 9). Using rule-based behavioral coding, sequence-oriented uptake analysis, student-level feature aggregation, and qualitative case comparison, the study examined both broad interaction patterns and specific forms of student follow-up after chatbot support. Broad quantitative comparisons showed descriptive but mostly non-significant differences across groups. The increase group showed the highest mean rate of code revision after assistant support (0.053), while the decrease group showed the highest mean rate of verification after support (0.131), with verification after support approaching but not reaching conventional significance (p = .057). Refined uptake analysis further suggested that broad productive follow-up was too general to separate groups clearly: increase and decrease students showed similar transformative uptake, while the decrease group showed the highest checking/help-seeking uptake. Three qualitative cases illustrated revision-oriented, checking-oriented, and mixed or fragmented interaction trajectories. These findings suggest that learning-gain groups were not clearly distinguished by overall chatbot use alone; rather, the form of student uptake after chatbot support may be a more informative descriptive lens. The study is framed as exploratory and does not make causal claims.

ABSTRACT. Component-based reconstructive learning has already been used in several domains as a practical design in which learners reconstruct a meaningful structure from given components. In this paper, reconstruction does not mean simply assembling arbitrary components. Rather, it refers to an activity in which an intended reference structure is decomposed into components, and learners reconstruct it from a component set designed to be necessary and sufficient for reconstructing that structure within the task. This constrained component set reduces the burden of generating components from scratch while requiring learners to interpret component meanings, judge semantic relations, and organize them into a coherent structure. However, because components are presented in advance, this approach has remained insufficiently theorized as a form of constructive learning. This paper provides a theoretical foundation for positioning component-based reconstruction as a legitimate constructive learning design by distinguishing free externalization from reconstructive externalization. Free externalization involves expressing already segmented understanding, forming and adjusting understanding through externalization, and leaving some target content unexternalized. Reconstructive externalization, by contrast, externally supports part of component generation and focuses learners’ activity on bridging the semantic gap between components and structure. By organizing this relation through the concepts of the duality of externalization, cyclic difficulty, semantic gap, and unreconstructed content, this paper clarifies why reconstruction-based learning should be recognized not as a simplified substitute for constructive learning, but as an educationally useful design that supports learners’ participation in structure construction and strengthens its theoretical intelligibility and adoptability. Keywords: Reconstruction-based learning, Concept map, Semantic gap