ABSTRACT. As Large Language Models (LLMs) are increasingly adopted for automated grading and feedback in higher education, their structured outputs, including multi-dimensional rubric scores, detailed feedback comments, and improvement suggestions, create an appearance of thorough analytic evaluation. This study examines whether these outputs deliver what they appear to offer. Using the JorGPT dataset of 3,041 student responses to 50 open-ended computer science questions, scored by both human instructors and three commercial LLMs, we identify three systematic discrepancies between the apparent and actual quality of LLM-generated grading and feedback. The sub-dimension scores are highly correlated (r = 0.82–0.99, VIF up to 45), providing redundant rather than independent diagnostic information. The textual feedback rarely detects student misconceptions (5–7% vs. 15–31% for teachers), functioning as a coverage checklist rather than a diagnostic instrument. The feedback tone remains uniformly positive regardless of response quality, lacking the severity modulation observed in human feedback. Additionally, grading accuracy varies significantly by knowledge domain, with procedural topics most reliable. These findings provide empirically grounded guidance on which aspects of LLM-generated grading and feedback can be relied upon and which require continued human oversight.

ABSTRACT. As generative AI systems become embedded in learning ecologies, learners are increasingly exposed to AI-generated content whose epistemic status, whether it represents verified fact, plausible inference, or fabrication, is difficult to discern. This paper examines the educational implications of two recently characterized AI failure modes: possibility-to-actuality hallucination (P2A hallucination), wherein AI systems treat mathematically possible but evidentially unsupported claims as factual, and distributed confabulation, wherein human–AI interaction creates a feedback loop that materializes and reinforces unsupported inferences. We argue that these phenomena pose a distinctive challenge for AI literacy education that current curricula are not equipped to address, because they focus on teaching learners that AI can hallucinate but not how to recognize the subtle, plausible hallucinations that are most resistant to detection. Drawing on constructivist learning theory, epistemic cognition research, and AI literacy frameworks, we propose a Pedagogical Framework for Evidential Reasoning with AI (PF-ERA) that integrates modal reasoning competencies into existing literacy curricula. We present five classroom-ready pedagogical scenarios, map these onto Bloom’s revised taxonomy, and analyze how P2A hallucination propagates through learning ecologies when AI-generated content circulates among learners, teachers, and institutional knowledge systems. Our framework addresses a critical gap in reimagining learning ecologies for the age of intelligent technologies: developing learners’ capacity to evaluate not just whether AI output is true, but how well-supported it is by available evidence.

ABSTRACT. The boundaries of reliable autonomy in agentic large language models (LLMs) remain insufficiently characterized, particularly in complex, document-intensive institutional workflows that require sustained reasoning across multiple information sources. This study investigates those boundaries within the context of educational administration by examining an Individual Program of Study (IPS) planning pipeline as a real-world case study. Using a rubric-based evaluation framework, the study assesses model performance across a multi-stage document analysis workflow, measuring accuracy, consistency, and reasoning quality at each stage of the process. The analysis identifies points at which autonomous performance degrades and traces the underlying reasoning behaviors associated with these failures. Findings indicate that retrieval augmentation alone does not ensure reliable execution of complex educational administrative tasks, especially when workflows require policy interpretation, cross-document synthesis, and judgment under ambiguity. Based on these results, the paper proposes a framework for AI delegation in curriculum planning that distinguishes tasks suitable for autonomous execution from those requiring human-in-the-loop (HITL) oversight. The framework offers practical guidance for institutions seeking to integrate agentic AI systems into administrative decision-making while maintaining reliability, accountability, and procedural integrity.

ABSTRACT. In classroom lectures, the instructor's body might block blackboard content. This occlusion degrades learning in cognitive load theory and downstream AI tools that process lecture videos. A survey about AI educational systems reveals that none of them involves occlusion handling. To bridge this gap, we present an innovative approach named Borrat to detect and remove teacher occlusions and restore blackboard content using computer vision techniques. On a real-world classroom dataset with teacher occlusions, our model achieves 97.99% performance, with only 4.5% latency overhead. We also evaluated our technique on two real-world MIT linear algebra lectures, and restored content regions yielding 8.9% more OCR characters. Moreover, we generated flashcards and multiple-choice questions from extracted text and classified them using Bloom's Revised Taxonomy. Results show that restored videos can produce more high-quality questions in higher-order thinking. We discuss the temporal split-attention problem in education based on cognitive load theory and how our technique addresses it, as well as its implications and applications in concrete use cases.

ABSTRACT. The seductive details effect occurs when learners’ attention is diverted by interesting but irrelevant information, potentially impairing comprehension and retention. While prior research in text-based instruction shows that seductive details placed early hinder learner performance more than when placed later, it is unclear how such placement affects engagement in video lectures. This experimental study investigated how the presence and placement of seductive details relate to learner performance and engagement across four video conditions: seductive details-only, no seductive details condition, and two mixed versions with seductive details appearing in the first or the second half of the video. The results indicated no statistically significant differences in test performance across conditions. However, placement was associated with differences in engagement measures. Specifically, the mixed conditions were associated with reduced task-unrelated thoughts and decreased difficulty disengaging from one’s own thoughts compared to the other conditions. These findings suggest that incorporating seductive details in the first half of a video may help support learner engagement without compromising learning outcomes, offering implications for the design of online instructional materials.

ABSTRACT. Team-based capstone projects in software engineering degrees often use project management tools, which can aid students in managing their projects and provide them with skills transferrable to tools they will encounter in the industry. However, commercial project management tools sometimes present issues in education contexts, such as useful features being paywalled or missing, and customization for the education context being limited because they are closed-source. Further, research dependent on data extracted from project management tools can also be limited by data commercial tools limiting what data can be exported. A solution for these issues is a purpose-built, open-source project management tool. In this paper, we present ScrumBoard, an open-source project management tool developed for both the education context and to support research. We discuss design decisions in ScrumBoard that are beneficial for conducting research with recommendations for researchers developing their own tools, and provide case studies of research projects ScrumBoard has facilitated to show practical applications of ScrumBoard’s research-oriented design. Through these case studies, we demonstrate that a purpose-built, open-source tool such as ScrumBoard, which is equipped with extensive logging and feature flags, and is customizable to our needs, enables research that would otherwise be difficult or impossible to undertake with commercial project management tools.

ABSTRACT. Generative AI is increasingly used to create educational content, yet most tools continue to position instructors as users of preconfigured systems rather than authors of learning environments. This paper examines how instructors engage in AI-mediated authoring of interactive learning simulators through vibe coding—a conversational, intent-driven practice in which pedagogical goals are translated into executable simulation logic through natural-language interaction, without formal programming. Focusing on a teacher-authored market sizing simulator developed and deployed in undergraduate management courses, we analyze authoring artifacts, parameter structures, and iterative design traces to identify how instructors formalize pedagogical judgment through model constraints, variable relationships, and scenario logic. Three design insights emerge: (1) pedagogical intent is expressed primarily through constraint-setting rather than content specification; (2) vibe coding functions as a reflective practice that externalizes tacit teaching judgment; and (3) instructor authoring redistributes teaching labor upstream into pre-class design, reshaping classroom orchestration. We argue that vibe coding represents an emergent form of AI-mediated instructional design that lowers technical barriers while preserving instructor agency over the structure of learning environments. The paper contributes design considerations for faculty-facing AI authoring tools and situates vibe coding within broader discussions of human–AI collaboration in educational technology.

ABSTRACT. Establishing a foundation of artificial intelligence (AI) literacy for all has become a critical necessity in societies where AI has fully permeated the workplace. It is vital to demonstrate practices that foster AI literacy, guiding participants from an introductory understanding to a deeper understanding of machine learning (ML). This study evaluates a 39-hour foundational AI course offered within a university programme for executive professionals, examining its impact on their understanding of ML. Employing a mixed-methods approach, data from 32 participants were analysed through a 27-item ML test and reflective writing tasks, administered both pre- and post-course, to assess changes in their ML understanding. Quantitative data were analysed using the Wilcoxon signed-rank test, while qualitative reflections were subjected to in-depth textual analysis. The quantitative results revealed a statistically significant improvement in ML test scores, with a large effect size (p < .001, r = .72). Qualitative findings indicated a notable progression in participants' ML literacy, evolving from a general understanding of ML concepts to a structured technical taxonomy of ML. Specifically, participants advanced from recognising basic concepts such as supervised, unsupervised, and reinforcement learning to demonstrating a more sophisticated understanding of ML and its practical applications in real-world contexts.

ABSTRACT. As AI education expands across the Global South, a critical question remains underexplored: what do children already believe about AI before formal instruction begins, and how do those beliefs shape their readiness to learn? This paper examines the perceptions and emerging explanatory models of AI held by 120 government school students in Grades 5–8 across four municipal and Zilla Parishad schools in Pune, India. Using a purpose-designed Marathi-language survey combining quantitative items, scenario-based reasoning tasks, and drawing prompts, we identify three recurring sense-making orientations: device-bound, anthropomorphic or cognitive, and process-oriented. We show that device-bound perceptions generate a hardware misconception - the belief that AI cannot be learned without devices - functioning as a conceptual access barrier in low-resource school contexts. Crucially, nearly half of hardware-bound students simultaneously demonstrated latent machine-learning intuitions, suggesting conceptual capacity is present but pedagogically unreached. An AI recognition task further reveals a YouTube paradox: the platform through which 73.5% of students first learned about AI is the one fewest can recognise as AI. We argue that equitable AI education must begin by mapping sense-making orientations learners already bring, requiring neither devices nor internet.

ABSTRACT. Conversational skills go beyond the ability to decipher text and use words together in a proper grammatical structure to comprehend and convey an intended message. It entails pragmatic competence or the ability to use language effectively in a social context, manage turn-taking, adapt to topic shifts, adhere to cultural norms, and repair communication breakdowns to achieve communicative goal. Pragmatic competence training can be facilitated through role-playing where learners practice language and navigate contextual nuances by acting out social scenarios. In this paper, we describe the design and implementation of Socius, a conversational agent architecture that leverages a Large Language Model to provide pedagogical scaffolding. Socius offers a platform where learners can engage in role-playing conversation with an AI Partner and receive immediate, rubric-based feedback from an AI Mentor. The system was evaluated over a 5-day usage period by young adults aged 18-24 years old (N=23). Quantitative results from a 5-point Likert scale indicate high user acceptance across three metrics: performance (M = 4.542), humanity (M = 4.468), and affect (M = 4.519). Qualitative feedback indicates that students perceived the agent as increasingly human-like over time, with emotional engagement correlating closely with the perceived quality of the AI’s role-playing and feedback. The AI Mentor scored participants’ pragmatic competency as very good at logical intent, proficient in achieving communicative goals, but struggle with politeness and emotional tone. These findings demonstrate that LLM-driven dual-role agents can provide adaptive, context-sensitive environments essential for mastering pragmatic competence while balancing universal pragmatic rules with the informal reality of local dialects and personal identity.

ABSTRACT. The increasing integration of Generative Artificial Intelligence (GenAI) in language education presents unprecedented opportunities for personalized feedback. However, existing research has primarily evaluated the static quality of AI-generated feedback, leaving the underlying, dynamic interactional processes underexamined. This study employs Lag Sequential Analysis (LSA) to investigate the behavioral patterns within a five-week corpus of learner-AI interactions (80 turns) collected during an undergraduate English debate training program. Interaction logs between an intermediate-level English as Foreign Language (EFL) learner and a large language model debate coach were coded according to learner behaviors (initiation and advancement) and AI feedback functions (evaluation, diagnosis, scaffolding, and elicitation). LSA revealed seven statistically significant behavioral transitions that form a continuous dialogic cycle. These pathways highlight a three-stage process: learner-driven triggering, AI pedagogical processing, and a learner response loop. The findings indicate that learner-AI interaction is organized around recurring dialogic feedback cycles, in which evaluation, diagnosis, scaffolding, and elicitation collectively support continued interaction. Crucially, targeted AI elicitation served as a necessary catalyst for learner advancement, effectively pushing the learner toward higher-order metacognitive negotiation. These findings suggest that GenAI functions not merely as a feedback provider, but as a dynamic co-regulatory partner that facilitates reflective and iterative learning processes.

ABSTRACT. This study examines how Japanese English as a Foreign Language (EFL) learners engage with machine translation (MT) in academic writing through an eight-week instructional intervention incorporating pre-editing and post-editing activities. Using a profile-informed explanatory sequential mixed-methods design, the study identifies learner profiles based on perceptions and usage patterns and explores their developmental trajectories over time. Quantitative analysis revealed five distinct learner profiles, while qualitative findings showed that MT use develops differently depending on initial profiles. Specifically, some learners remain constrained by anxiety and limited strategies, whereas others develop strategic and self-regulated use through active engagement. The findings suggest that effective AI use does not emerge automatically from access to tools but requires instructional support, particularly through scaffolding and collaborative learning. The study proposes a developmental model of MT use and highlights its implications for fostering AI literacy in language learning. In addition, the findings suggest that the skills developed through MT use, such as critical evaluation and revision of generated output, constitute foundational competencies for engaging with generative AI, suggesting the need for instructional design that supports learners’ adaptive use of generative AI tools.

ABSTRACT. Large language model tutoring systems can generate student-state summaries that appear in teacher dashboards, influence problem selection, and shape feedback. Such a summary is a form of educational hallucination when it derives from the problem statement or a repeated annotation template rather than from what the student actually said. We present inverse student simulation (ISS), a trustworthy-AI audit framework for this risk. ISS estimates a structured student-state record from tutoring dialogue and then checks whether that record is recoverable from dialogue evidence and whether it changes a forward simulator’s predictions. The record, Z, encodes mastery probabilities for 30 knowledge components (KCs), 70 misconception-code probabilities, and four exploratory metacognitive scalars. Using MathDial dialogues, we find that broad human labeling is reliable but model recovery is uneven. Two coders labeled all 394 silver-labeled test dialogues and reached substantial to almost-perfect agreement (κ = 0.701–0.847). The fine-tuned ISS inverter shows only weak within-dialogue KC ordering (Spearman ρ = 0.274; top-3 overlap = 40.7%), and a label-free prefix heuristic achieves lower calibration error than the full-dialogue ISS model (Brier 0.028 vs. 0.038). Misconception scores are dominated by a degenerate four-label pattern shared across all 394 test dialogues; the inverter emits a near-constant misconception vector (across-dialogue σ ≈ 0.002) that a zero-parameter label-prior baseline matches without reading any dialogue. Metacognitive recovery is unreliable, with most scalar correlations near zero. Forward replay shows the full-context model predicts responses with little dependence on Z (mean NLL difference = 0.006), while Z-only probes confirm Z carries signal when context is removed. These are negative results, not a working model: under degenerate silver supervision the inverter collapses to a label prior rather than recovering student state, so ISS's contribution is the reusable, model-agnostic audit framework and this quantified failure-mode diagnosis. Labels must pass audit before reaching dashboards, problem selection, or feedback.

ABSTRACT. Although the use of artificial intelligence (AI)-generated feedback (FB) is promising for enhancing learning, evidence showing how its internal composition affects student learning remains scarce, in particular in elementary education. Most previous studies treat FB as a categorical variable instead of decomposing its characteristics as continuous predictors. Longitudinal lag effects—the ways through which FB in one lesson influences reflection in the next—remain underexplored. This study investigates how AI-generated FB composition, which is operationalized through Hattie and Timperley’s FB model [including the levels of Process (FP), Self-regulation (FSR), and Self (FS)], longitudinally affects the quality [Structure of the Observed Learning Outcome (SOLO) scores] and quantity (character count) of social studies reflections by elementary students. We performed a 6-month design-based research field study (August 2025–February 2026) in a Japanese public elementary school (N = 14 students, 691 observations, 64 class sessions). The FB system, implemented using GPT-4.1 mini and GPT-5.1, was iteratively refined through three phases of the design (Types 1–3). The numbers of FP, FSR, and FS aspects per FB text were coded as continuous variables and entered into linear mixed models as lagged predictors. First, FP aspects in previous FB exhibited consistent positive correlations showing subsequent reflection quality and quantity (ρ = .102–.168). Second, FSR significantly predicted character count (β = 5.44, p = .009), but FS negatively predicted SOLO scores (β = −0.162, p = .022; marginal after false discovery rate correction, q = 0.066). Third, growth trajectories accelerated in Phase 2, when FP and FSR increased and FS decreased in prompt redesign (Phase 2 × class session: β = 0.051, p < .001 for SOLO); this elevated level was sustained through Phase 3. Because FB Type and time period are confounded and Phase 3 incorporated a model upgrade, strict causal attribution to FB redesign requires additional controlled studies. The decomposition of AI FB into continuous aspect-level variables produces differential dose–response relationships across FB levels, as well as supporting design guidelines that minimize Self-level praise and enrich Process and Self-Regulation content.

ABSTRACT. With the rapid advancement of Large Language Models (LLMs), their integration into educational contexts has accelerated, including tools that generate personalized stories to support kanji (Chinese characters) learning for elementary school students. However, the educational appropriateness and linguistic naturalness of these AI-generated narratives require rigorous validation, as conventional automated metrics often fail to capture subtle textual nuances. This paper investigates the feasibility of utilizing LLMs to automatically detect and classify specific types of textual "oddness" in LLM-generated stories, with two objectives: (1) evaluating how well LLM-based classifiers reproduce human-defined oddness judgments. (2) identifying ambiguities in annotation criteria through analysis of LLM-generated rationales. We defined a typology of seven oddness categories organized into three hierarchical levels—local, contextual, and educational oddness—along with a "Natural" category. A dataset of 480 LLM-generated stories was annotated by human raters. and inter-annotator agreement was analyzed using Cohen's kappa. Moderate agreement was obtained for the Natural category, whereas agreement for individual labels was at or below chance level. indicating that the boundaries between labels remain ambiguous. LLM-based classification was evaluated using a Multi-Label Confusion Matrix (MLCM); the highest F1 score in the eight-label setting was 0.237, whereas binary detection of oddness presence achieved an F1 score of 0.69. These findings suggest that annotation ambiguity, rather than model capability alone, constitutes the primary barrier to reliable automatic evaluation of AI-generated educational narratives. This study provides foundational insights for developing benchmark tools to evaluate the quality of AI-generated educational content.

ABSTRACT. This study focuses on learning by quiz-creation as a learning strategy for fostering information literacy competence in Japanese education. It examines the extent to which LLM-based agent-assessment is reflected in learners’ quiz corrections. We propose a learning support system that enables learners to correct their own quizzes based on LLM-generated assessment results. An experimental study was conducted with 21 high school students in a quiz-creation activity on the Monte Carlo method. The results showed that 52.1% of valid assessment feedback items were reflected in learners’ corrections, suggesting that agent-assessment can support quiz correction. However, the reflection rates were polarized, indicating that the feedback was not equally effective for all learners. Feedback on typographical errors and terminology consistency was more likely to be corrected, whereas omission errors and feedback on the clarity of quiz statements were less likely to be reflected. These findings suggest that agent-assessment is effective for supporting formal improvements, but feedback presentation should be improved to reduce learners’ cognitive load during correction.

ABSTRACT. Generative AI is increasingly positioned as a tool for personalised learning, yet less is known about how young learners participate in shaping AI-mediated support in classroom practice. Building on prior conceptual work on agentic personalisation (AP), this study examines how AP was enacted and constrained in Korean elementary science classrooms. AP is defined as a learner-authored and teacher-mediated process in which students shape learning goals, orchestrate AI-supported pathways, evaluate AI-generated responses, and transform AI-generated outputs into their own learning artefacts. Drawing on an ethnographically informed qualitative case study, the main analysis focused on 41 Grade 6 students who used GenAI during lessons on plant structure and function. Data included student worksheets, AI dialogue records, reflections, final written explanations, and teacher-generated observation logs. Eight focal cases were selected through maximum variation sampling to capture contrasting patterns of AI engagement. Findings show that some students enacted AP beyond prompting by questioning the factual reliability of AI responses or transforming AI-generated explanations into meaningful analogies. However, productive prompting did not always lead to output ownership; some students relied on AI wording, repeated vague questions, or drifted off-task. The study contributes to AI in education research by showing that AP is a learner-authored, teacher-mediated, and unevenly enacted process rather than a form of system-delivered adaptation.

ABSTRACT. This study investigated Aotearoa New Zealand (NZ) teachers' perceptions and integration of Large Language Models (LLMs) in teaching, with particular attention to place-based education (PBE) contexts. Data were collected from 182 in-service NZ teachers via a survey combining scales and open-ended questions. Using an embedded mixed methods design, quantitative data were analysed through hierarchical regression and mediation analysis; qualitative data analysis through deductive thematic analysis guided by quantitative findings. Perceived usefulness emerged as the dominant predictor of AI integration, suggesting that attitude and belief function as proximal cognitive mediators through which utility perceptions are translated into adoption decisions. Despite moderately positive attitudes, trust was lower than midpoint, shaped primarily by professional culture, teaching level, and discipline identity. PBE experience did not predict AI integration significantly, suggesting that PBE educators neither simply resist nor embrace AI. Teachers navigated the tension between LLM and PBE through selective adoption. Using AI for generic procedural tasks while maintaining PBE practices for culturally grounded content involving Māori history including local pūrākau (narratives) and oral histories. These findings suggest mainstream LLMs' structural limitations for place-based and Indigenous educational contexts where knowledge is relational, oral, and community-embedded, highlighting the need for culturally responsive AI tools.

ABSTRACT. Agentic artificial intelligence (AI) systems are autonomous, goal-directed tools capable of decomposing tasks, generating intermediate reasoning steps, and dynamically adapting within learning environments. These capabilities are increasingly shaping how learners engage in problem solving and knowledge construction. While such systems can improve task completion and provide structured guidance, their influence on learning processes remains insufficiently understood. In particular, limited attention has been given to how overreliance on agentic AI systems may allow learners to bypass knowledge construction processes, such as problem analysis, reasoning, and strategy development that typically occur during learning, and how this may influence learning outcomes, including retention and knowledge transfer. This study develops a conceptual framework to examine whether agentic AI functions as a cognitive amplifier that supports comprehension, or alternatively contributes to a comprehension illusion in which task performance improves without gains in durable learning outcomes. The framework integrates two complementary theoretical perspectives: Cognitive Load Theory and Self-Regulated Learning Theory. From a cognitive load perspective, AI-generated reasoning steps may reduce extraneous cognitive load by simplifying problem-solving processes. However, overreliance on such support may reduce germane cognitive effort required for schema construction, retention, and knowledge transfer. From a self-regulated learning perspective, AI-generated solutions may reduce learners’ engagement in planning, monitoring, and evaluation. The design will examine differences in cognitive load, self-regulated learning processes, and learning outcomes, including retention and transfer.

ABSTRACT. While expectations for Evidence-Based Education continue to grow, significant challenges remain in the systematic extraction of evidence from Real-World Educational Data (RWED). Among various methodologies, causal discovery has emerged as a promising focus for the automated extraction of evidence from such data. However, conventional algorithms often face practical limitations, such as the necessity for large-scale datasets to achieve stability and an inherent vulnerability to unobserved confounding factors, which are common in educational settings. To address these limitations, this study proposes a novel methodology that integrates the vast relational knowledge of Large Language Models (LLMs) into the causal discovery process. Specifically, we incorporate causal relationships by an LLM as prior knowledge into LiNGAM-MMI, which is a causal discovery algorithm distinguished by its robustness against confounding and its effectiveness with small-scale datasets. We conducted comprehensive evaluation experiments using both standard causal analysis benchmarks and RWED. The results demonstrate that the proposed method achieves superior accuracy compared to existing baseline techniques on benchmark data and exhibits high compatibility when applied to complex educational datasets. These findings suggest that the integration of LLM-based prior knowledge with LiNGAM-MMI is a highly promising approach for the extraction of Real-World Evidence (RWE) from Real-World Data (RWD). The outcomes of this research are expected to facilitate the routine and continuous collection of RWE directly from educational practice, thereby accelerating the practical and social adoption of evidence-based education.

ABSTRACT. Handwriting logs, including pen pressure, offer less intrusive data because they can be collected during writing activities, without requiring separate monitoring devices. Although pen pressure has been linked to engagement-related states, most studies have examined it in controlled settings, leaving its relevance in naturalistic classroom contexts largely underexplored. Yet, real classroom situations may shape learning engagement differently, especially for students with special needs, whose participation may depend on their individual characteristics and classroom support. Therefore, it is necessary to examine whether pen pressure can provide meaningful insight into students’ engagement during regular learning activities. This study explores pen pressure data from six special needs students, collected on a digital learning platform over a one-month period of music therapy lessons. Two aspects were investigated: how pen pressure differs across student attention levels, and how this information was interpreted by the teacher in relation to students’ characteristics. Pen pressure intensity, variability, and within-task change patterns were analyzed across students in high and low attention groups based on teacher ratings. Results showed that pressure variability distinguished attention groups more clearly than pressure intensity. Fluctuating patterns were observed exclusively among low attention students, particularly during tasks with less teacher guidance. Reflexive thematic analysis of the teacher interview suggested that pen pressure patterns may reflect students’ affective states, cognitive effort, and transitions in attention. Overall, when combined with teacher interpretation, pen pressure data may serve as behavioral indicators of students’ learning engagement during handwriting-based tasks. These findings suggest their potential as a foundation for developing automated learning monitoring systems in special needs education.

ABSTRACT. Generative artificial intelligence (GenAI) has increasingly entered design education, yet its educational value should not be understood merely as the production of visually attractive design images. In sustainable design education, novice learners often need to connect abstract environmental concepts, such as passive ventilation, daylighting, material reduction, thermal buffering, and biophilic integration, with observable spatial and material cues. This short paper examines how prompt-controlled AI-generated images can function as teacher-mediated visual scaffolds for sustainable design learning. Rather than evaluating image generation only as a technical representation problem, the study focuses on how learners interpret unannotated visual materials and how pedagogical cue clarity may support conceptual understanding. The proposed framework includes feature classification, visual cue operationalization, prompt formulation, image screening, and learner evaluation. Fifty interior-level low-carbon design features were organized into five categories and generated under controlled scene and rendering conditions using Stable Diffusion. Eighty-four design-background participants evaluated six randomly assigned images without textual labels. The preliminary results show that objective recognition accuracy was moderate, whereas perceived instructional usefulness was relatively positive. Visual realism alone was not significantly associated with objective recognition, suggesting that pedagogical cue clarity may be more important than photorealistic quality and that cognitive load should be considered when unannotated AI-generated images are used with novice learners. The paper contributes to computers in education by reframing text-to-image GenAI as a teacher-mediated visual scaffolding process rather than a purely creative production process. It also provides methodological groundwork for future X+AI cross-domain design research, especially craft-vocabulary-guided interior scene generation and mixed-reality co-creation learning.

ABSTRACT. The increasing availability of urban environmental data creates new opportunities for residents to access information about air quality and environmental noise. However, access to data does not necessarily translate into the ability to interpret and use it in everyday decision-making. This study examines citizen-facing environmental dashboards as a communicative and potentially educational layer between complex environmental monitoring systems and urban residents. A scoping review methodology was employed. The review included 52 open-access publications and was designed as a descriptive and interpretive evidence-mapping study. The analysis identified seven functional types of digital environmental interfaces. Based on the synthesis, a seven-level model of environmental communication is proposed. The findings indicate a gradual shift from dashboards focused primarily on presenting environmental indicators towards systems that support spatial and temporal interpretation, forecasting, risk assessment, scenario analysis, recommendations, and natural-language interaction. The findings may support urban authorities, environmental agencies, public health organizations, and civic technology designers in developing more pedagogically informed digital interventions that facilitate residents’ understanding of environmental data and support action to reduce pollution.

ABSTRACT. The advancement in artificial intelligence (AI) gives rise to large language models (LLMs) capable of producing text. LLM is coherent and appropriate to the context provided by input prompts. These emerging technologies are used in educational settings. However, there remains a lack of research on how human evaluations of storytelling compare with those made by LLMs, particularly character design and visual components. This study examines the effectiveness of LLMs in evaluating children’s storytelling abilities and explores the impact of paper-based character assembly on enhancing their creativity and storytelling skills. Specifically, it assesses the performance of open- and closed-source LLMs in a Traditional Chinese educational context, aiming to understand how these LLMs compare with human evaluations. The study involves 11 primary school students who participate in storytelling sessions before and after engaging in paper-based character assembly activities. We utilize a triangulation methodology that includes evaluations of story performance, time consumption, and analysis of character assembly. Both human raters and LLMs, including the novel Deepseek-r1-distill and other models, such as GPT-3.5-turbo and Llama33-70b, are employed to evaluate the narratives. The results demonstrate that the Deepseek-r1-distill model displays high stability and strong correlation with human ratings, outperforming closed source models, showing lower consistency and weaker correlations. Additionally, engaging students in the assembly of paper-based characters noticeably enhances their creativity and storytelling abilities, as evidenced by the development of more complex and personalized narratives. The findings highlight the potential of integrating LLMs into educational settings to enhance children’s storytelling skills. The study also reveals the beneficial effects of active learning approaches in fostering creativity and engagement in storytelling. These insights provide valuable directions for educators seeking to incorporate innovative assessment and learning strategies.

ABSTRACT. Natural Language Processing (NLP) techniques are rapidly integrating into daily life in the form of mobile and web applications, mainly for educational purposes. Educational sectors are making use of this to improve language learning methods. However, there is a lack of such applications for Tamil language and efforts in that domain remain low, despite a huge Tamil population globally. This paper explores the implementation of different NLP techniques based on various Tamil datasets and their integration into a gamified Tamil language learning platform, targeted at Primary School students to build their foundation in a meaningful and engaging manner. Our results, obtained through various models and methodologies, highlight the high potential of this platform along with key limitations of current Tamil datasets. This study provides practical insights on the creation of a gamified Tamil language learning platform using NLP techniques, contributing toward more interactive and AI-assisted Tamil educational applications for younger learners.

ABSTRACT. This study investigates how learner-AI interaction patterns differ according to learners’ critical thinking skills in generative AI-assisted EFL reading and how these interactions relate to summary writing performance. An experiment was conducted with 40 adult EFL learners. Participants first completed a critical thinking assessment and received prompt guidelines. They then interacted with ChatGPT while reading and comprehending an English passage, followed by a summary writing task completed without AI assistance. Learner-AI interaction patterns were analyzed using Epistemic Network Analysis (ENA), and the effects of interaction on task performance were examined through moderated regression analysis. The results revealed that learners with higher critical thinking skills generated significantly more Critique and Deep Reasoning prompts than those with lower critical thinking skills. ENA further showed that the interaction patterns of the high critical thinking group were characterized by stronger connections involving Deep Reasoning, whereas the low critical thinking groups’ interactions were centered on Comprehension-related prompts. In addition, Feedback Request emerged as a significant predictor of summary writing performance regardless of learners’ critical thinking levels. These findings suggest that critical thinking is reflected in the ways learners interact with generative AI during reading activities. Furthermore, interaction behaviors that encourage monitoring and reflection, such as Feedback Request, may contribute to improved reading comprehension outcomes. The study highlights the importance of incorporating reflecting interaction strategies into the design of AI-assisted foreign language learning environments.

ABSTRACT. Productive failure pedagogy intentionally places learners in challenging problem‑solving situations prior to formal instruction, making emotional experience and regulation central to learning. In a secondary school classroom study (N = 159), I examined which emotions students experience during the initial collaborative problem-solving phase, how they naturally regulate emotions, and how experimentally guiding savoring or dampening of the most strongly felt emotion affects regulation uptake and learning. Results show that students’ emotional experiences are functionally organized rather than valence‑based, with epistemic emotions like interest and confusion playing central roles. Most students do not naturally engage in emotion‑aware problem-solving, instead relying on emotion‑minimizing or neglecting strategies. Although dampening negative emotions was easier to interpret than savoring, intuitive regulation strategies did not translate into delayed learning benefits. Critically, posttest learning showed a significant interaction between emotional valence and regulation condition: dampening supported learning when emotions were negative, whereas no regulation was the most effective when emotions were positive. Findings demonstrate that intuitive emotion regulation is not necessarily productive during failure‑driven learning and highlight the need for emotion‑aware scaffolds aligning regulation with epistemic goals.

ABSTRACT. Collaborative learning depends not only on how much students speak, but also on how they respond to, build on, and influence one another's ideas. However, feedback for group discussions often relies on quantitative activity indicators, such as the number of utterances or posts, which may overlook students who speak less frequently but play important interactional roles. This paper proposes an LLM-based network analysis method for visualizing interaction-based student contributions in group discussions. The method segments discussion transcripts into topic units, extracts explicit relations among utterances with a large language model, and constructs a directed student network in which edges represent response relations such as agreement, disagreement, clarification requests, and evidence-based critiques. The HITS algorithm is then applied to calculate two indicators: Social Impact, which represents the extent to which a student's utterances elicit responses from others, and Responsivity, which represents the extent to which a student responds to others' utterances. We applied the method to six discussions conducted by 12 undergraduate and graduate students and compared the results with expert ratings, self-ratings, peer ratings, LLM zero-shot ratings, and utterance counts. The results showed that LLM zero-shot ratings corresponded more closely with expert ratings, whereas the proposed method showed lower correlations with utterance counts. A questionnaire further showed that LLM zero-shot feedback was easier for students to understand, while the proposed network-based feedback helped some students notice response relations among members that were difficult to recognize during discussion. These findings suggest that the proposed method should be positioned not as a replacement for expert or LLM holistic ratings, but as complementary structural feedback for supporting students' reflection on collaborative interactions.

ABSTRACT. Peer feedback is widely used in educational contexts, but students often find it difficult to interpret feedback and use it to improve their work. The recent use of artificial intelligence (AI) offers new ways to support this process, yet little is known about how different forms of AI support influence students’ behaviours when acting upon peer feedback. Drawing on the feedback literacy framework, this exploratory study examined students’ sequential patterns of acting upon peer feedback and the timing of judgement making under three conditions: No-AI, AI-assisted, and AI-directed. In the AI-assisted condition, a chatbot provided clarification and prompts without directly rewriting text. In the AI-directed condition, a chatbot generated revised text that students could adopt, edit, or reject. Thirty-nine graduate students were randomly assigned to the three conditions in an argumentative writing revision task. Screen recordings yielded 223 behavioural episodes from 34 valid participants, which were coded into three dimensions: attending to feedback, making judgements, and taking action. Sequential pattern analysis suggested different behavioural patterns across conditions. Students in the No-AI condition mainly engaged in repeated self-authored rewriting after attending to peer feedback. Students in the AI-assisted condition showed more varied behaviours, often comparing peer feedback, AI responses, and their own drafts before integrating or rewriting suggestions. Students in the AI-directed condition frequently adopt AI-generated revisions, followed by later evaluation, editing, or rejection. The distribution of specific feedback literacy codes differed significantly across conditions, χ² = 43.79, p < .001, and descriptively judgement making occurred earlier in the AI-supported conditions. These findings add process-level evidence to feedback literacy research by showing how students work with peer comments, AI-generated input, and their own task goals when acting upon peer feedback. They also suggest the need to design AI support that guides students to interpret, compare, and justify their responses to feedback.

ABSTRACT. When a student uses an adaptive tutoring system, the system must decide whether the student has learned a skill well enough to advance or should keep practicing. This decision is controlled by a mastery threshold, a numerical cutoff applied to the student model's estimate of mastery: when the estimate reaches the cutoff, the system advances the student; otherwise practice continues. As intelligent learning environments increasingly swap classical student models for neural ones, many systems reuse the same threshold, assuming that the same cutoff has the same practical meaning across model types. This paper tests that assumption by evaluating six knowledge tracing (KT) models, computational methods that estimate student mastery from practice records, across four public educational datasets. Using a shared, reproducible evaluation framework, we compare post-mastery hit rate, the fraction of correct responses in the first five practice items after advancement; learner coverage, the fraction of students who ever reach the threshold; practice burden, the number of attempts required before advancement; and a transparent instructional-utility score across 14 thresholds from 0.50 to 0.99. Bayesian Knowledge Tracing is relatively insensitive to threshold increases, while neural models show sharp trade-offs between reported post-mastery performance and learner coverage. Higher thresholds are positively associated with reported post-mastery performance, negligibly for BKT and moderately for neural models, but for neural models this association is about half explained by which students are selected to advance rather than by improved readiness for all who practiced. Reusing a Bayesian Knowledge Tracing-calibrated threshold with a neural model produces the largest instructional utility loss among all transfer directions tested. Stricter thresholds also do not close performance gaps between students with stronger and weaker prior records and can widen disparities in which students reach advancement. The paper provides an architecture-aware calibration framework and practical, equity-conscious guidance for selecting or recalibrating mastery thresholds when tutoring systems change student models or instructional goals.

ABSTRACT. With the rapid expansion of online education, there is an increasing demand for video content in which teachers provide explanations alongside lecture slides. However, the production of such videos requires high costs and intensive labor. In this paper, we propose a novel method for automatically generating lecture videos with a teacher avatar performing semantic gestures synchronized with the spoken content. The system requires only lecture slides and a single full-body picture as input. This research contributes to the field of online education by lowering the cost to lecture video production and enhancing the social presence of instructors due to the automated generation of expressive, context-aware movements. The proposed system first utilizes a Large Language Model (LLM) to generate a lecture script based on the provided slides. This script then is used as the basis for generating skeletal motions that performs contextually appropriate gestures. Finally, by generating the avatar using the skeletal motion and the single full-body picture, the system produces a video in which the teacher avatar delivers a lecture with natural, expressive movements. Subjective evaluations with 20 university students were conducted to assess the effectiveness of the proposed method. The results demonstrated that the lecture video with the teacher avatar significantly improved "Social Presence," "Understandability," and "Desire to Watch" compared to videos without an avatar, suggesting the pedagogical value of semantic gestures in automated lecture video production.

ABSTRACT. As generative artificial intelligence (GenAI) becomes increasingly embedded in classroom practice, understanding how teachers and AI jointly construct teaching decisions has emerged as a critical yet underexplored challenge. Existing research largely focuses on outcomes or tool effectiveness, offering limited insight into the underlying decision-making processes. To address this gap, this study developed a process-sensitive multimodal diagnostic framework for teacher–AI collaborative decision-making. Using a mixed inductive–consensus design, Phase A applied procedural grounded theory to 30 authentic classroom interaction recordings (1,282 coded instances) to generate candidate indicators. Phase B refined the framework through a two-round Delphi study with 18 experts, and Phase C employed the Analytic Hierarchy Process to derive normalized indicator weights. The results yielded a three-stage decision model—Information Sharing, Strategy Internalization and Generation, and Collaborative Construction—comprising three primary and nine secondary indicators. Weighting results highlighted Collaborative Construction as the most influential stage, with Feedback Regulation identified as critical. The framework provides a replicable foundation for diagnosing, comparing, and optimizing teacher–AI decision processes in instructional contexts.

ABSTRACT. Generative AI can provide learners with explanations and answers easily, but such output may be accepted without sufficient meaning construction. This study proposes necessitation of meaning construction as a design principle for learning with AI-generated output. Based on this idea, we designed a two-stage task in which learners first construct a concept map from an AI-generated explanatory text and then reconstruct it from decomposed components. The task is intended to prompt learners to externalize and reexamine conceptual relationships in AI-generated explanatory texts. To explore this design, we conducted an exploratory study to examine question generation across phases, revisions of conceptual relationships in maps, and learners’ subjective evaluations. The results showed that the construction and reconstruction phases were perceived to contribute more to deeper understanding than the initial reading phase, and that both learners’ questions and concept maps changed across phases. These findings provide initial evidence that the proposed task may promote active reinterpretation of AI-generated explanations.

ABSTRACT. Intelligent tutoring systems increasingly use LLMs to classify dialogue moves and inform adaptive scaffolding decisions. However, aggregate accuracy may hide errors on dialogue turns that are important for pedagogical decision-making. We propose VALID-ITS, an audit framework that evaluates LLM dialogue coders across four dimensions: subgroup and metric audit, rewrite stability audit, context and boundary audit, and scaffolding risk audit. We evaluated LLMs and supervised classifiers across three public corpora: TalkMoves, CIMA, and MathDial. The results show that aggregate accuracy and macro-F1 can diverge sharply, errors can concentrate at speaker-turn boundaries, and some pedagogically important error pairs can produce high wrong-scaffold rates. In TalkMoves, prompt for reasoning to revoicing reached a WSR of 1.000 for one coder variant. These findings suggest that LLM dialogue coders should not be deployed in ITS based only on aggregate accuracy.

ABSTRACT. Artificial Intelligence (AI) Technologies (e.g., large language models) make it possible to deliver immediate, adaptive feedback at scale, offering a potential substitute for the costly authoring effort required for human-designed feedback. Yet it remains unclear whether different structures of AI feedback yield learning gains comparable to human feedback, and how these effects vary across learners who differ in prior knowledge and feedback literacy. We conducted a randomized study with 302 college students assigned to one of three feedback conditions: (a) human feedback (business-as-usual), (b) AI one-layer feedback (provide correct answer paired with elaboration and suggestions in all attempts; non-layered), or (c) AI two-layer feedback (provide encouragement, hints, and reflective prompts in the first attempt, correct answer in a later attempt; layered). Learning was assessed through pre- and post-tests, and participants additionally reported their feedback literacy (the understandings, capacities, and dispositions needed to make sense of and act on feedback), and perceptions of the benefits of the feedback they received. Results show that, first, both AI feedback conditions produced post-test gains comparable to human feedback overall, supporting the viability of AI feedback as a scalable alternative. Second, feedback literacy (FL) but not prior knowledge moderated the effect: low-FL learners benefited more from AI feedback (both one-layer and two-layer) than from human feedback, while high-FL learners performed comparably across conditions. Low-FL students were more likely to perceive AI feedback as promoting self-regulated learning benefits (i.e., positive affect and motivation as well as self-reflection and improvement). These findings suggest that AI feedback of different structures can match the learning value of human feedback and may be particularly valuable for learners who enter the task with lower feedback literacy, which also informs further AI system design to adapt to learner characteristics.

ABSTRACT. SQL-Tutor records how students submit SQL queries, receive feedback, and revise after constraint violations. This short paper uses these traces to examine whether students with similar prior knowledge worked with the tutor in different ways. The analytic sample included 83 students who completed both pre- and post-tests and had complete behavioral log measures. Seven standardized features were derived from the logs: adjusted learning-curve slope, learning-curve consistency, adjusted constraint recovery, help-seeking intensity, progression efficiency, improving-problem rate, and average logged-in time per attempt. Gaussian mixture modeling identified three profiles: Slow/Low Progress Learners, Efficient Progress Learners, and Persistent Self-Repair Learners. The profiles did not significantly differ in pre-test scores, but they differed in error trajectories, repair behavior, pace, progress, and learning efficiency. Post-test scores and normalized learning gains followed the expected direction but were not significantly different after correction. The analysis therefore treats repair and progress traces as learner-modelling evidence in their own right, not only as predictors of test scores. The study shows how domain-model traces from a constraint-based tutor can support interpretable profiles for future adaptive scaffolding.

ABSTRACT. Designing well-calibrated assessments in large-scale online degree programmes involves a structured, multi-stage review process in which instructors explicitly decide the question type, mark allocation, and cognitive difficulty level of each item. We argue that this institutional process encodes item difficulty and discrimination, as formalised by the Two-Parameter Logistic (2PL) Item Response Theory (IRT) model, into the observable structural features of the questions, and we test this hypothesis empirically using eleven consecutive terms of a Statistics course from an online degree programme at a tier-1 Indian university (181 questions, 4,649 students). Ground-truth item parameters are estimated via an Expectation-Maximisation(EM) algorithm applied to historical student response data. We evaluate five regression models across all 64 subsets of six-question feature groups (320 combinations) under a strict longitudinal expanding-window protocol. The principal finding confirms the hypothesis: question type and mark value are the most informative features for predicting both difficulty and discrimination. A simple conditional mean estimator conditioned solely on these two features achieves a mean combined Root Mean Squared Error (RMSE) of 0.9784, closely approaching the best machine learning estimator (Random Forest (RF), 0.9126), and requires no model fitting beyond a lookup table of historical averages. This has a serious implication for assessment designers as it provides a more explainable and robust mental model while designing assessments for this course in future. As a downstream application, predicted item parameters are combined with prerequisite-course-based student ability estimates via the 2PL response function to produce student quiz score predictions before any current-term response is collected, achieving a score RMSE of 33.755 against an oracle lower bound of 28.350. These results provide both a practically deployable tool for pre-assessment calibration and a data-driven explanation of why structural design choices reliably encode item psychometric properties.

ABSTRACT. This study presents the first large-scale, taxonomy-based analysis of micro-credential (MC) metadata characteristics across a national higher education ecosystem, analyzing 233 digital badges stratified-sampled from a national registry of 2,858 publicly available MCs across nine institutional segments. Using the IACET Badging Taxonomy as an international benchmark, two independent coders evaluated each MC against three criteria: Granularity, Self-containment, and Alignment. The results revealed three features of the Japanese MC ecosystem. First, MCs skewed heavily toward IACET Levels 3–4 (course completion), with only 13.7% (n = 32) reaching Level 5-6, suggesting that MCs function as digitized certificates rather than as evidence of competence. Second, a significant characteristic imbalance was observed (Friedman χ²(2) = 154.25, p < .001) on MCs. Third—and most critically—we uncover a "standardization paradox": Data Science MCs exhibited the highest alignment (Dunn's test, p < .001) yet a lower granularity than Specialized MCs, as voluntary benchmark alignment replaced granular learning outcomes with generic institutional boilerplate. We argue that MCs must evolve toward learner-centric, competence-driven design, and we offer concrete metadata-level guidelines for instructional designers and digital credentialing platforms within the ICCE research community.

ABSTRACT. Educational institutions increasingly collect multi-source student data, including attendance records, academic performance, daily mood self-reports using a weather metaphor, and health room usage. These data can support learning analytics and evidence-informed educational decision-making, but their use is constrained by privacy risks, fragmented data systems, and limited technical expertise among practitioners. Recent large language models (LLMs) offer promising natural language interfaces for data exploration; however, LLM-assisted analytics may expose sensitive records, enable small-group re-identification, or support inappropriate profiling if connected directly to student-level databases. This paper presents a privacy-preserving LLM-assisted learning analytics platform designed for secure educational environments. The system integrates municipal school data through a four-layer architecture consisting of raw, clean, feature, and Model Context Protocol (MCP) layers. The raw and clean layers store original and normalized student-level data, but they are never exposed to the LLM or end users. Instead, the LLM accesses only aggregated, k-anonymity-protected feature tables through whitelisted MCP tools. The platform combines architectural separation, schema guards, read-only access, k-anonymity suppression, and audit logging to protect against unauthorized access, small-group disclosure, and unsafe automated operations. The prototype was implemented using PostgreSQL, Python-based ETL pipelines, MCP tools, and a locally deployed LLM environment. We demonstrate how the system enables natural language queries about attendance, well-being, and learning outcomes while enforcing privacy-by-design principles. This work contributes a reproducible design pattern for responsible LLM-assisted learning analytics in sensitive educational contexts.

ABSTRACT. Generative artificial intelligence (GAI) is increasingly reshaping creative learning practices, yet limited research has examined how pre-service teachers (PSTs) experience and adopt GAI in multimodal pedagogical design. This qualitative study investigates PSTs’ perceptions of GAI-supported digital storytelling (DST) through the lens of Rogers’ Diffusion of Innovation theory. Seventy-nine PSTs participated in an 18-week DST project in which they used tools such as ChatGPT, Midjourney, Runway, and CapCut to develop scripts, storyboards, images, animations, and final digital stories. Data were collected from reflective journals, focus group interviews, and stimulated recall interviews, and were analyzed thematically according to five perceived innovation attributes: relative advantage, compatibility, complexity, trialability, and observability. The findings show that PSTs perceived GAI as a productive co-creation partner that accelerated script development, visual generation, animation prototyping, and collaborative decision-making. At the same time, they encountered challenges related to prompt formulation, output inconsistency, cross-tool coordination, technical instability, and the pedagogical appropriateness of GAI-generated content. Participants responded to these challenges by refining prompts, combining multiple tools, post-editing outputs, negotiating design choices within groups, and iteratively aligning GAI-generated materials with their narrative and educational goals. The study suggests that GAI adoption in teacher education is not a linear acceptance process but a situated process of human-GAI negotiation, creative adaptation, and pedagogical judgment. These findings contribute to understanding how intelligent technologies can be integrated into teacher education to support creative, collaborative, and reflective learning ecologies.

ABSTRACT. Proactive Teacher Professional Development (TPD) requires reliable environments for pedagogical rehearsal; however, orchestrating complex group dynamics in Computer-Supported Collaborative Learning (CSCL) imposes severe cognitive burdens on educators. Existing Large Language Model (LLM) simulations fail to provide such reliable rehearsals due to an “alignment tax”, where models auto-correct seeded student misconceptions, causing “logic hallucinations.” To bridge this fidelity gap, we propose a Data-Grounded Cognitive (DGC) architecture that anchors multi-agent behaviors in authentic, multimodal student profiles (cognitive anchors, metacognitive confidence, and social personality traits). We validated the system through a Tri-Level Alignment Study, benchmarking in-silico logs against empirical ground-truth from a remote CSCL session. Results demonstrate high individual and group alignment in foundational tasks; however, complex tasks reveal a decline in process alignment, empirically exposing the LLM alignment tax. Crucially, Lag Sequential Analysis confirms profound structural isomorphism between the simulated socio-metacognitive flow and authentic discourse hegemony. The DGC architecture effectively provides a high-fidelity “Human-on-the-loop” (HOTL) sandbox, advancing evidence-based TPD.

ABSTRACT. AI guidance systems for vocational learners are being deployed at scale across India and comparable contexts, yet their design is driven by what the technology can do rather than what learners need, feel, and will accept. This paper addresses that gap through a co-design study with 28 young people enrolled in Industrial and Vocational Training Institutes across eight Indian states. Participants took part in a six-hour structured workshop comprising three activities: mapping the questions they carry and the emotions those questions produce, tracing their journeys across information sources to understand how trust forms and breaks, and co-designing the behavioural boundaries they consider non-negotiable for an AI guidance system. The study was conducted to inform the design of AskAbhi, an AI chatbot developed by Quest Alliance for the MyQuest learning application. Three findings emerge. (1) Learners bring emotionally charged, high-stakes questions, not information requests. Six thematic domains emerged; career and employment questions accounted for 28% of all questions, and AI and technological uncertainty for 17%. Confusion was the most frequent emotional response across all domains. Fear characterised 50% of AI-related questions, and 67% of decision ambiguity questions were associated with confusion. (2) Learners do not trust AI as a primary authority. Help-seeking followed sequential, multi-source paths: AI was used alongside teachers, family, peers, and search engines, and trust formed only when multiple sources aligned. Breakdown moments arose when learners needed help choosing between options they already had, not when they lacked information. (3) When learners defined what they wanted from AI, they led with what it must never do. Across 183 co-designed boundary statements, the dominant categories were answer style, accuracy, and privacy, each framed as prohibitions. Structured clarity, accuracy, and privacy emerged as preconditions for trust, not desirable features. Six design principles derived from these findings offer a grounded framework for AI guidance systems in high-pressure, low-resource vocational education contexts.

ABSTRACT. Our prior work proposed a four-dimension framework for the qualitative analysis of programming-video annotations. This paper concerns the first dimension, Note Type, which classifies annotations by surface form and has since served as an upstream classifier for cognition-aware feedback, dashboards, and deviation detection, yet has never been independently validated. We test its reproducibility on 1,471 annotations from three years of an undergraduate Python course, isolating the 87 system-tagged Questions and comparing six perspectives on the remaining 1,384 notes. We compare expert annotation, a rule-based codebook, unsupervised clustering with multilingual sentence embeddings and character n-grams, a supervised classifier trained by five-fold cross-validation on expert labels, and large language models under zero-shot and few-shot prompting conditions. The LLM conditions are run with both Claude Sonnet 4.5 and GPT-4o. Expert agreement reaches Cohen's κ = 0.83. The rule-based codebook reaches κ = 0.68 against expert consensus. Unsupervised clustering and LLM induction do not recover the taxonomy, instead defaulting to Python-topic structure rather than surface form. With expert calibration, two operationally distinct methods reach expert-comparable agreement: a supervised classifier over the codebook's own signals reaches κ = 0.78, statistically above the codebook, while a few-shot Claude classifier with 30 expert examples reaches κ = 0.73, not statistically distinguishable from the codebook. The supervised classifier and the few-shot LLM are also not statistically distinguishable from each other. These results show that the Note Type dimension is reproducible by calibrated automated methods at different expert-label budgets, but that this dimension is imposed by the codebook and expert examples rather than emerging naturally from the corpus. Whether the result extends to the framework's three deeper dimensions remains open.

ABSTRACT. Large Language Models (LLMs) have created new opportunities for scalable programming support, but many existing tools prioritize direct coding assistance and remain external to institutional learning environments. These limitations make it difficult to provide context-aware feedback, cultivate Self-Regulated Learning (SRL), and give instructors visibility into how students interact with AI assistance. Building on our prior conceptual design and interview study, this paper presents the implementation, public release, and preliminary evaluation of CodeRunner Agent, an open-source LLM-powered SRL scaffolding system integrated with the Moodle CodeRunner plugin. The system combines an embedded student support panel, SRL-oriented AI scaffolding, automatic SRL phase detection, code guardrails, and an instructor dashboard for real-time learning analytics. The student interface supports four operational phases: Planning, Program Creation, Error Correction, and Self-Reflection. We conducted a preliminary evaluation with 12 students and 8 instructors using interaction logs, questionnaires, and open-ended feedback. Log analysis showed that students used the system across multiple SRL phases, especially for error correction. Students and instructors reported generally positive perceptions of the system’s usability, context-aware support, and hint-based feedback. These findings provide design implications for embedding SRL-oriented LLM support into existing programming learning environments.

ABSTRACT. Generative AI can readily provide working code, creating a tension in introductory programming: students need timely support, yet direct solutions may reduce opportunities for reasoning. To address this, we propose example-based scaffolding, in which GenAI generates scaffold examples that share a target task’s underlying reasoning pattern but differ in surface context, supporting analogical transfer while reducing the risk of direct copying. Specifically, we contribute a two-dimensional taxonomy of scaffold examples with operationalized criteria for identifying effective scaffolds, and CodeExemplar, a prototype that integrates scaffold example generation into a web-based coding environment with automated test-case grading. We report findings from a formative evaluation consisting of a classroom pilot with 30 students and a prototype evaluation with 13 students and 4 instructors. Students perceived scaffolded examples as useful for making progress without revealing the target solutions and preferred them over other forms of support. Instructors also viewed the approach as promising for transfer-focused practice. We discuss design implications, limitations, and directions for future studies.

ABSTRACT. Generative artificial intelligence (GenAI) is increasingly shaping learners’ cognitive, behavioral, and emotional engagement in education. This umbrella review examined what literature reviews have been conducted and what results have been reported on GenAI and emotional engagement. Fourteen eligible review articles published between 2023 and 2026 were identified and analyzed. Findings show that relevant evidence is distributed across three strands: reviews on multidimensional learner engagement, reviews on adjacent emotional constructs, and reviews on emotion-sensitive AI capabilities. Emotional engagement was most directly discussed as one dimension of learner engagement, but its indicators varied across contexts, including interest, enjoyment, confidence, value, relevance, reduced anxiety. Existing reviews suggest that GenAI may support emotional engagement-related outcomes through feedback, personalization, conversational interaction, and task support. However, gaps remain in construct clarity, mechanism synthesis, and measurement.

ABSTRACT. Many students’ blind trust in generative AI (GenAI) tools has raised concerns about over-reliance and cognitive offloading in AI-assisted learning. This study distinguishes between two forms of trust: destructive trust and critical trust. Destructive trust refers to trusting AI mainly because it is convenient, efficient, or appears to work well, which may lead to over-reliance and negative learning consequences. Critical trust, in contrast, refers to trusting AI through careful evaluation, awareness of its limitations, and a balance between confidence and questioning. In the present study, we developed a Socratic Questioning AI Agent that defers direct answer generation and engages students through six types of Socratic questions, with the aim of exploring whether this approach can promote critical trust in AI. We examined how 40 postgraduate students used and responded to the designed agent. Self-reported questionnaire results showed an increasing trend in trust, a significant decrease in perceived ease of use, and no significant change in reliance. Further analysis of students’ reflective writing revealed active cognitive engagement during their use of the agent, suggesting the development of critical rather than destructive trust. These findings suggest that pedagogical design can reshape the typical relationship between trust and reliance in AI-assisted learning.

ABSTRACT. Learning Analytics Dashboards (LADs) hold promise for supporting learner self-monitoring and data-driven decision-making, yet learners frequently neglect dashboard access and associated data behaviors. Gamification has been proposed as a means to address this challenge, but how gamification elements relate to intended behaviors at the level of fine-grained interaction remains largely unexplored. This study analyzes 55,603 interaction logs generated by 110 K8 students over a 15-day gamified Quantified Self activity, alongside pre-survey data on data use efficacy and awareness. Using Transition Network Analysis (TNA) and mixed-effects logistic regression, we investigate how gamification elements are associated with expected behaviors on the dashboard and which engagement factors within a gamified dashboard predict daily dashboard access. TNA results show that the gamification node achieved the highest betweenness centrality, functioning as a recurring hub through which students navigated toward learning records, data analysis, and inquiry recording behaviors. Regression results identify mission completion rate and consecutive access days as positive predictors of next-day access. In contrast, cumulative dashboard access days showed a negative association, suggesting that novelty-driven engagement in the early stages of the program gave way to gradual disengagement over time. These findings offer actionable insights for gamification design oriented toward sustaining intended learning behaviors in student-facing LADs.

ABSTRACT. Alzheimer’s disease is a neurological disorder that deteriorates a person’s brain which impairs cognitive skills such as thinking, reasoning, and memory. Despite the Philippines’ aging and affected population, awareness on the disease is lackluster as misconceptions are still present in the media. This aim of the study is to bring more awareness regarding the effects of Alzheimer’s disease on the patient and their interpersonal relationships through a narrative-driven video game titled The Long Goodbye. The mechanics are based on research of the cognitive impacts of Alzheimer’s disease, while the story is influenced by conducted interviews with Filipino caregivers. As such, this study explores how game design can serve as a medium for education and emotional engagement regarding Alzheimer’s disease. From a round of playtesting, it was observed that the narration in the game helped deepen the understanding of the players on the social issue. However, there is still room for improvement in terms of specifying how the mechanics the players experience relate to the issue rather than merely just an aesthetic choice.

ABSTRACT. The current paper aims to investigating the role of incorporating a humanistic approach into virtual reality (VR)-based humanities lessons by means of gamification. Gamification is described within the paper by analyzing three literary masterpieces, such as *To Kill a Mockingbird*, *The Kite Runner*, and *The Stranger*. These examples helped to determine the most important elements of gamified learning, such as narrative plot, role-playing, branching pathway, and trial-and-error approach. Gamified learning contributes to the development of a personalized approach to learning, empathy, and reflection; thus, it promotes a more efficient learning process. In addition, it was proven that VR-based learning in the humanities can be considered a gamified system contributing to the development of individuality, error handling, and identity formation. The described techniques promote the achievement of Sustainable Development Goals (SDG 4 and SDG 10). This framework shows how VR-based humanities education can be viewed as gamified education, where storytelling, acting, and problem-solving are used to encourage students and help them learn deeply.

ABSTRACT. The increasing use of generative artificial intelligence (GAI) in education raises questions about how AI tools can be pedagogically orchestrated to support meaningful learning rather than automate task production. In teacher education, pre-service teachers must not only use AI but also critically revise and refine instructional designs. This study examined a peer-assessment-enhanced GPT-supported condition (PA-GPT) to investigate its effects on pre-service teachers’ instructional design learning. A quasi-experimental study involving 82 pre-service teachers compared PA-GPT (n = 43) with a comparison ChatGPT-supported condition (C-GPT, n = 39). Results indicated that PA-GPT was associated with significantly higher performance in completeness, innovation, and interaction, as well as higher critical thinking tendency and self-efficacy, and lower cognitive load. Interview findings further suggested that peer feedback and comparison prompted learners to reconsider design rationales and compare alternative design choices. The findings highlight the added value of structured peer assessment within a generative AI- and concept map-supported instructional design environment.

ABSTRACT. Generative artificial intelligence (GenAI) has been increasingly used as a learning tool in higher education. However, students’ final answers or quiz scores may not fully show how they engage with course content. This study examined how structured and unstructured GenAI-supported reflection made students’ learning performance, cognitive engagement, and self-review processes visible in an undergraduate course. A total of 28 students participated in a four-week learning activity using TronClass with GenAI-supported functions. The control group completed open-ended reflection tasks with GenAI support, whereas the experimental group completed a Bloom-based structured worksheet and a GenAI self-review form. Data included weekly AI-generated quiz scores, worksheet responses, and GenAI self-review responses. Worksheet responses were coded according to Bloom’s cognitive levels and grouped into surface-, middle-, and deep level cognition. Self-review responses were coded for conceptual awareness, error awareness, self-monitoring, self-regulation, and transfer of learning. The results showed that the experimental group had higher weekly quiz scores across the four weeks. Both groups produced evidence of surface-, middle-, and deep level cognition, but the experimental group showed more consistent distributions across cognitive levels. The GenAI self-review responses further showed students’ awareness of concepts, errors, unclear knowledge, learning strategy changes, and transfer to real-world scenarios. These findings suggest that Bloom-based structured worksheet and GenAI self-review forms can provide teachers with observable information about students’ learning processes when GenAI is used as a learning tool.

ABSTRACT. As AI rapidly permeates K–12 education, the transition from basic tool acquisition to effective pedagogical integration remains a significant challenge. This mixed-methods study evaluates the effectiveness of a targeted teacher development (TD) workshop designed to enhance the subject-specific TPACK and metacognitive capacity of 116 in-service primary English teachers in Hong Kong. The four-hour intervention utilised a customised, curriculum-aligned AI chatbot specifically designed to scaffold the teaching of writing about location and direction. Data were collected through pre- and post-workshop surveys and written qualitative reflections. Wilcoxon signed-rank tests revealed statistically significant improvements (p < .001) across all TPACK sub-domains—CK, PCK, TCK, and TPACK—and all sub-domains of metacognitive capacity. Thematic analysis of teacher reflections corroborated these findings, highlighting empowerment to adopt new pedagogical approaches and a recognised shift from traditional instruction towards fostering student metacognition and autonomy. This research demonstrates that customised, subject-specific TD is essential for moving beyond generic AI use. Future research should investigate the long-term effects of such interventions on classroom practice and student outcomes.

ABSTRACT. This study examines the collaborative relationship between Learning Analytics (LA) researchers and in-service teachers in the field of learning analytics research. Although this relationship has long been assumed to exist, it has never been formally defined. Although collaboration between researchers and teachers is widely recognized as essential in LA practice research, its implementation has relied more on implicit expectations and ad hoc arrangements than on formal agreements. To address this issue, we conducted a Policy Delphi study consisting of three rounds of structured dialogues with four LA researchers and six K-12 (kindergarten through high school) in-service teachers. Based on 10 research topics covering the entire cycle of LA practice research, we analyzed points of agreement and disagreement between the two stakeholder groups. The results revealed that while collaboration was generally recognized as desirable, there were significant initial discrepancies in expectations regarding roles, authority, and workload. While the Policy Delphi process contributed to strong consensus on four topics, it exacerbated conflicts on others, particularly those related to methodological authority and the right to withdraw from the research. Based on these findings, we derived three design principles for LA practice research policy: the clear articulation of conditions for collaboration, the differentiation of structured roles, and the institutionalization of structurally contentious issues. These principles offer a practical foundation for formalizing researcher-teacher relationships in LA practice research and highlight the need for concrete policy instruments rather than reliance on goodwill alone.

ABSTRACT. Modern industrial education needs new digital tools to teach practical skills. While Information and Communication Technology (ICT) is now widely used in regular classes, vocational high schools and technical colleges still face a serious problem. Skilled engineers use many senses—sight, hearing, touch, and the feeling of force—to judge their work. This "sensory tacit knowledge" is very hard to teach through videos or words alone. This study proposes a new e-learning material, called "E-learning by Doing," that uses a sensory transmission device to transfer this knowledge to learners. As a first step, we focused on the manual feed operation of a lathe, where keeping a stable feed rate determines surface quality. The proposed material delivers an expert's feed-rate motion to learners as direct physical stimuli, helping them calibrate their own speed sense without conscious imitation. To evaluate the material, 27 beginners participated in an experiment as a control group (12, video only) or an experimental group (15, video plus device), and performed face and taper machining on a real lathe. In face machining, the experimental group showed improved surface profile quality, especially in Pa (ANCOVA: Pa p = 0.025, η² = 0.192, large effect). The correlation analysis showed that the control group followed a single stabilization path (ρ=-0.713, p=0.009), whereas the experimental group achieved good quality through different motion strategies.

ABSTRACT. As a representative form of smart learning environment, the smart classroom provides robust technological support for classroom interaction. However, existing research has predominantly focused on developing generic analytical frameworks while paying insufficient attention to the distinctive pedagogical demands of specific subjects, making it difficult to yield discipline-specific instructional implications. Drawing on social interaction theory, this study proposes four teaching strategies for promoting classroom interaction in primary school Chinese language instruction within smart classroom environments: constructing interactive instructional resources, cultivating a harmonious classroom climate, implementing heterogeneous collaborative learning, and providing timely evaluation and feedback. A quasi-experimental study was conducted with two fifth-grade classes from a primary school. Classroom videos were coded using an interaction coding scheme tailored to the Chinese language discipline and were subsequently analyzed through behavioral frequency analysis and Lag Sequential Analysis (LSA). The results indicate that the quality of classroom interaction in the experimental class was significantly higher than that in the control class, suggesting that the proposed strategies can effectively enhance classroom interaction in primary school Chinese language teaching.

ABSTRACT. Classroom video offers rich insights into teaching and learning processes but remains underutilized due to challenges in scalability, accessibility, and data preprocessing. This paper presents COREference, an AI-enabled approach for transforming large-scale classroom video data into a curated corpus for research and teacher professional learning. The study addresses a critical bottleneck in classroom video corpus construction: the reliability of audio quality and transcript generation. We propose a two-layer architecture that separates automated data curation from human–AI interaction. The first layer implements an AI-supported preprocessing pipeline that integrates multi-feature acoustic analysis (RMS energy, spectral centroid, and spectral flatness) for audio quality assessment, alongside domain-specific automatic speech recognition (ASR) model selection for transcript generation. The second layer implements a large language model-based framework (Qwen3-4B-Thinking-2507) to support semantic retrieval and prompt-guided pedagogical code inference using aligned video frames, transcripts, and expert-defined pedagogical code schemes. Current findings show that audio quality is a key determinant of transcription accuracy, with multi-feature filtering effectively improving downstream ASR performance. Comparative evaluation of ASR models across subject domains further reveals that transcription quality varies by language and classroom conditions, motivating subject-specific model selection. Preliminary implementation of the Layer 2 framework demonstrates the feasibility of using LLMs to support structured pedagogical coding and semantic exploration of classroom data, while reducing reliance on manual annotation processes. The proposed approach supports improved corpus reliability and usability, highlighting the potential of AI-supported preprocessing for scalable educational research infrastructures.

ABSTRACT. This study proposes the RDIF_R tree, which extends RDIF_R, a statistic used for detecting uniform differential item functioning (DIF) within the Residual-Based DIF framework, into a tree-based structure to identify DIF that may threaten test fairness and measurement validity. The RDIF_R tree explores possible split points based on examinee covariates without requiring groups to be defined in advance, thereby enabling the exploratory identification of subgroups in which DIF occurs and interaction structures among covariates. In this study, simulation data were generated based on the three-parameter logistic model (3PLM), and the detection performance of the RDIF_R tree was examined under conditions varying in the proportion of DIF items, DIF magnitude, Sample size and covariate structure. The results showed that the unpurified RDIF_R exhibited a substantial increase in Type I error rate as the proportion of DIF items increased, whereas the RDIF_R tree controlled the Type I error rate stably at a level comparable to the purified RDIF_R. In addition, the RDIF_R tree maintained high power and demonstrated its usefulness in exploring subgroups exhibiting DIF and interaction structures when compared with the Item-Focused tree. These findings suggest that the RDIF_R tree can be used as a method for exploring DIF associated with group characteristics that are not specified in advance, while preserving the computational efficiency and interpretability of the conventional RDIF_R.

ABSTRACT. Collecting large and diverse datasets of student robotics programs is inherently difficult because activities depend on physical hardware, classroom settings, and teacher-designed tasks, limiting the data available for AI-supported learning analytics and automated program analysis. This paper presents a semantics-aware data augmentation framework for block-based educational robotics programs. We represent LEGO SPIKE Prime projects as mutable block graphs and define a set of controlled, semantics-preserving transformations over visible, pedagogically meaningful constructs including ports, units, actuator commands, variables, expressions, conditionals, and loops. Semantic equivalence is grounded in an abstract robot model, enabling transformations to preserve program behavior independently of concrete hardware configuration. An empirical evaluation on approximately 40,000 student programs shows that augmentation strength is easily controllable, and that generated variants increase both structural and parameter-level diversity. In a downstream task-classification experiment, augmented training data improved accuracy for each evaluated classifier family. These results demonstrate that semantics-preserving transformation-based augmentation is a practical strategy for expanding educational robotics program datasets and supporting AI-based program analysis in a domain where large-scale data collection remains a persistent challenge.

ABSTRACT. he Chinese Proficiency Test (HSK) is one of the most widely used assessments in international Chinese education. This study examines whether a large language model (LLM) can support reliable item generation for HSK Level 4 reading comprehension. We built an integrated pipeline with three linked stages: deriving item-writing principles, generating items with prompt engineering, and validating the results with real test data. Drawing on the HSK syllabus, the national grading standard (GF 0025-2021), a review of prior research, and interviews with three subject-matter experts, we distilled 35 item-writing principles and turned them into a structured prompt. We then ran three studies using GPT-5.4. Study 1 generated items from a fixed corpus and asked seven experienced HSK item writers to review them. Study 2 used the model to screen reading texts and compared its decisions against an expert gold standard. Study 3 administered an automatically generated test form and an authentic HSK form to 120 learners and compared the two forms. The expert review showed good agreement (ICC = 0.82) and an average quality score of 2.32 out of 3. The model screened texts with an F1 score of 89.26 percent against the expert gold standard. In the field test, the generated form and the authentic form did not differ in a meaningful way (Cohen's d = 0.20), reached similar reliability (alpha = 0.78 versus 0.81), and showed comparable item difficulty and discrimination. About 78 percent of learners rated their overall test experience as positive. The findings suggest that an LLM-based pipeline can produce HSK Level 4 reading items of usable quality, while human review is still needed for distractor design and difficulty control.

ABSTRACT. Japan's growing multilingual school population has intensified the need for effective communication between schools and foreign guardians. While language barriers are widely recognized, a more fundamental challenge lies in the implicit sociocultural knowledge embedded in Japanese school communication—knowledge that conventional translation tools fail to convey. To address this gap, this study developed PAPAMAMA-TOMO, an AI-mediated multilingual support system designed to help foreign guardians interpret Japanese school culture through contextually grounded explanations. The system integrates a school-print knowledge base constructed from 2,754 school handouts, from which 1,792 implicit knowledge entries were extracted using NLP and LLM-assisted methods, combined with a Retrieval-Augmented Generation (RAG)-based conversational interface supporting multiple languages and image-based input. A preliminary pilot study was conducted with 26 foreign guardians, primarily Vietnamese-speaking (50%) and Nepali-speaking (42%), the majority of whom had lived in Japan for more than three years. Among the participants, 81% reported having used the system. Results indicated that approximately 70% of users experienced a reduced need to contact schools for clarification, and approximately 77% reported increased confidence in handling school-related communication. The explanation function was most valued, followed by translation support and situational guidance. These findings suggest that AI-supported sociocultural interpretation—beyond translation alone—holds significant promise for promoting educational inclusion in multicultural school communities.

ABSTRACT. Abstract: Generative AI writing assistants are widely used in universities, yet how different feedback designs shape students' self-regulated learning (SRL) behaviors during revision remains underexplored. This between-subjects experiment compared two AI feedback modes (direct output and Socratic questioning) as twenty Chinese EFL university students revised English argumentative essays. Using turn-level coding grounded in Pintrich's (2004) SRL framework (κ = 0.885) and Epistemic Network Analysis, we analyzed 165 student turns. Results showed that direct output engaged a Control–Reflection cluster while Socratic questioning engaged a Monitoring–Help cluster (Cohen's d = 1.51). Neither mode triggered complete P→M→C→R microcycles; knowledge-rule extraction was absent in both groups. The behavior–outcome relationship reversed across conditions: SRL behavior was unrelated to score gain in the direct-output group (ρ = −0.11) but positively associated in the Socratic group (ρ = +0.62). These results suggest that adaptive mixed-mode scaffolding, adjusting to students' current regulatory state, is needed.

ABSTRACT. In research organizations, members acquire experiential knowledge through daily research activities. However, such knowledge is often embedded in individual thought processes and remains tacit, making it difficult for subsequent members to understand how their predecessors interpreted problems, made decisions, revised artifacts, and derived reusable knowledge. We develop an experiential knowledge sharing platform based on thought experiences. Designed with reference to the Dual-Loop Model, the platform supports the externalization and accumulation of thought experiences, as well as the stepwise formalization and sharing of experiential knowledge derived from them. The platform treats evolving research artifacts as thought products and records their changes as versions and histories. It then uses accumulated thought experiences as cues for formalizing experiential knowledge, which is shared within the organization together with its formation context. A practice in a university laboratory suggested that the platform can support the construction and sharing of experiential knowledge based on individual thought experiences.

ABSTRACT. This study investigated how social comparison functioned in a student-facing dashboard during K-12 students’ summer vacation learning. Although social comparison is widely used in dashboards, it remains unclear when students engage in it during learning, whom they compare themselves with, and whether such comparison leads to metacognitive awareness. To address these issues, we analyzed the learning logs of 236 Japanese seventh- and eighth-grade students during their summer vacation assignment, including both logs from the e-book system and logs of their use of a dashboard that allowed them to compare their learning processes with those of others. Transition network analysis was applied to each learning actions, and memo data were coded using eight components of metacognitive awareness. The results showed that social comparison was embedded in the learning flow and was repeatedly used in an exploratory manner. Most comparisons were made with classmates, while prior-cohort comparisons were less frequent and often limited to all or high-performing seniors. Low-performing students were more likely than high-performing students to compare themselves with others at the start of learning, suggesting reliance on others’ progress as an external reference. However, only a small number of students wrote reflective memos, and no clear difference was found between high- and low-performing students in the amount or type of metacognitive awareness expressed. These findings suggest that social comparison can support learning engagement and decision making, but dashboards should also provide support that helps K-12 students interpret comparison information and turn it into metacognitive awareness.

ABSTRACT. As generative artificial intelligence (AI) becomes increasingly integrated into education, understanding its impact on students’ cognitive engagement during problem solving is critical. While behavioral and self-report measures provide some insight, they cannot capture the real-time neural dynamics of AI-supported learning. This study employed electroencephalography (EEG) to investigate cognitive engagement during problem-solving tasks with and without AI support. Using a within-subjects design, participants completed structured and ill-structured problems under human-only and AI-assisted conditions. EEG power spectral density in the frontal, occipital, and temporal regions was analyzed across preparation and solution phases. Linear mixed-effects models showed that AI-assisted problem solving was associated with reduced alpha and theta power, especially during the preparation phase. These results provide evidence of how AI modulates cognitive engagement and demonstrate the utility of EEG for examining the cognitive impact of generative AI in educational contexts.

ABSTRACT. Blended synchronous learning (BSL) is a highly practical instructional modality in higher education. However, it presents instructors with the challenge of monitoring the emotional engagement of online learners in real time. This paper presents an AI-powered Video Analytic Tool (AVAT), a system designed to monitor and analyze learner facial expressions during live Zoom sessions and provide instructors with real-time engagement data via a dashboard. The tool captures and classifies learning-relevant emotions at one-second intervals, distinguishing positive states (e.g., enjoyment, excitement) from negative ones (e.g., frustration, boredom, anxiety). A web-based prototype has been developed and integrated with the Zoom SDK and AWS cloud infrastructure. This paper presents the conceptual framework, system architecture, key features, and current prototype, and a plan for future validation and deployment in authentic BSL settings.