Personalized exercise recommendation aims to provide students with suitable learning materials based on their individual knowledge states and learning needs. However, existing recommendation systems still face two challenges: they often overlook the alignment between exercise difficulty and student proficiency, and they usually rely on a shared model that cannot adequately capture the heterogeneity among different schools. In this paper, we propose FedPER, a federated personalized exercise recommendation framework that integrates federated learning with a hierarchical attention mechanism to support personalized recommendations across heterogeneous educational environments. In FedPER, each school is treated as an independent client with a two-stage local model consisting of knowledge tracing and exercise recommendation. For knowledge tracing, a Graph Convolutional Network (GCN) is employed to model the relationships among knowledge concepts based on a Phi-coefficient correlation matrix, while a Gated Recurrent Unit (GRU) captures the dynamic changes in students' knowledge states over time. Based on the estimated knowledge states, the exercise recommendation module identifies students' weak concepts and selects exercises by jointly considering knowledge relevance and difficulty adaptation. On the server side, a hierarchical attention mechanism (FedHAM) is introduced to perform adaptive model aggregation by considering layer-wise differences between global and local parameters, thereby improving the model's ability to handle non-IID educational data. Experiments conducted on the ASSISTments 2009 dataset show that FedPER achieves competitive performance in terms of difficulty-adaptation accuracy, which measures how closely the difficulty of the recommended exercises matches each student's ability, and novelty compared with existing approaches under both federated and non-federated settings. The results demonstrate that FedPER can provide effective personalized exercise recommendations while adapting to the heterogeneous data distributions and learning characteristics of different school-level clients.
As generative artificial intelligence (AI) becomes increasingly visible in education, an important question concerns how AI-related knowledge structures are associated with children's scientific creativity. Prior studies have mainly examined this issue through performance outcomes. In contrast, the present study focused on knowledge representation. We examined whether children's scientific creativity was related to the similarity between their semantic organization and an AI-derived semantic reference. A total of 125 sixth-grade students completed a scientific questioning task and a word-association task. In parallel, we generated responses from 120 AI agents based on six contemporary large language models to construct an aggregated AI semantic reference. Individual semantic networks were derived from the word-association data, and representational similarity analysis (RSA) was used to estimate each child's alignment with the AI reference. The results showed a nonlinear relationship. Originality followed an inverted-U pattern, whereas the Composite Creativity Index (COM) showed a peak followed by a plateau. In both cases, the highest performance was observed at intermediate levels of alignment. As a contextual comparison, AI agents showed higher Originality but lower Flexibility than children, and their semantic networks were sparser and more modular. These findings indicate that scientific creativity was highest at intermediate levels of alignment with the AI-derived semantic reference. At higher levels of alignment, Originality declined, whereas Composite Creativity plateaued. The study provides a representational perspective on the nonlinear association between children's semantic organization and scientific creativity.
Accurately tracking students' evolving knowledge states is a key requirement for intelligent tutoring systems and adaptive learning platforms. Existing deep knowledge tracing models can capture temporal learning patterns, but their hidden states and gating mechanisms are often difficult to inspect. We propose Auto-TSCD, a focused extension that combines target-aware temporal representations with computational-tree search to discover structurally inspectable temporal cognitive diagnosis rules. Candidate residual rules are represented as computational trees over learned or model-derived inputs, including student and exercise embeddings, a historical-state representation, and target-aware temporal similarity. A probabilistic model-building strategy searches for rules that balance validation performance and structural simplicity, and the selected rule is reported as an explicit mathematical formula. Experiments on three real-world educational datasets show that Auto-TSCD obtains the highest mean AUC in the reported comparison, with modest gains of approximately 0.65–0.85 percentage points over the directly related TSCD and PMB-CAS models. On ASSISTments-09, PMB-CAS retains better ACC and RMSE. The explicit tree exposes which model-derived representations and operators enter the prediction, providing structural transparency; semantic and teacher-facing interpretability remain to be validated.
Monitoring classroom participation requires dependable, low-latency, private solutions that work within edge hardware constraints while maintaining meaningful learning connections. On-device inference is conceivable with edge-AI systems, but they rarely analyze input signal importance in real-world classroom scenarios and employ face orientation or gaze as crude proxies. The embodied cognition theory predicts that engagement should be represented in whole-body sensorimotor activity on specific tasks, not only in face configuration, but agentic classroom pipelines have not operationalized this requirement. In this article, we present agentic edge-AI engagement detection (AEED), a cross-disciplinary pipeline that leverages classroom deployment constraints to inform model design, data collection, and system development. In particular, AEED consists of BodyPosture (BP), a model trained on traditional camera-facing body-orientation cues, and FaceReaction (FR), a face model with labels derived from synchronized EEG, ECG, and task performance. In order to examine the impact of input design alongside architectural and platform characteristics, both models operate in parallel under the same tracking, hardware, and logging settings. The pipeline uses commodity edge hardware for interaction recording, dual-model inference, and anonymous student tracking. When camera geometry changes, BP's orientation-based signals invert, whereas FR's physiologically grounded signals correlate with learning outcomes in contexts similar to its training conditions. These results suggest that input design and label construction may matter more than architectural choice for deployment validity, although training-domain differences between the two models remain a partial confound. For agentic AI systems in authentic classrooms, these findings offer a reproducible framework for building embodied, per-student engagement trajectories.
Sign languages are visual-spatial languages that convey meaning through hand gestures, facial expressions, and body postures. While it plays a crucial role in fostering inclusion and communication between hearing and deaf communities, the integration of digital technologies in sign language learning, particularly for Taiwan Sign Language (TSL), remains limited. To address this gap, this study evaluated the usability and user experience of an integrated AR-AI platform, which is powered by a MediaPipe and Two-Stream I3D recognition framework to facilitate interactive TSL tasks. A total of 65 participants aged 18 and above, including hearing and deaf or hard-of-hearing (DHH) individuals, engaged with the system. Usability and user experience were evaluated using the Usability Metric for User Experience–Lite (UMUX-LITE) and the Standardized User Experience Percentile Rank Questionnaire (SUPR-Q), respectively. The results revealed that DHH participants rated the AR system's usability significantly higher than hearing participants (p < .001). Similarly, the AI-based interactive system received significantly more positive UX evaluations from the DHH group (p = .036). Furthermore, a positive correlation was observed between sign language proficiency and perceived usability, suggesting that existing linguistic mental models significantly influence interaction efficacy. These findings provide critical HCI insights into designing inclusive interfaces and highlight technical considerations for developing accessible TSL-based interactive technologies tailored to diverse user profiles.
Situational Judgment Tests (SJTs) are a well-established class of assessment instruments that embed measurement within realistic scenarios, eliciting context-driven responses that are less susceptible to social desirability bias and more effective at capturing applied judgment than traditional self-report scales. Despite these advantages, the widespread adoption of SJTs in educational assessment remains constrained by the high cost and labor intensity of manual item development. To address this scalability bottleneck, we propose AES (Agents for Educational SJTs), a multi-agent framework powered by large language models (LLMs) that automatically transforms existing Likert-scale items into psychometrically informed, context-rich SJTs. AES comprises three cooperative modules, Generator, Simulator, and Calibrator, that emulate the expert driven development pipeline by combining qualitative review from simulated domain experts with quantitative analysis of synthetic student responses, enabling iterative item refinement without human intervention. We instantiate AES in the domain of AI literacy, a complex and affectively rich competency that typifies the constructs for which SJTs are most beneficial. Evaluation on both human participant and LLM-simulated response datasets shows that AES-generated SJT items provide improved psychometric evidence relative to their Likert counterparts, while learners report greater engagement and perceived authenticity. These findings provide initial evidence that AES can serve as a scalable and psychometrically informed approach for supporting the development of scenario-based AI literacy assessments.
Agentic tutoring systems increasingly rely on large language models to generate real-time feedback, yet risky outputs such as answer leakage, contradiction, and hallucination can directly affect learning. Answer leakage is a central concern in this setting. It exposes problem solutions to students and can be verified through human annotation. This paper presents a pedagogical style-conditioned audit protocol for automated evaluators that detect answer leakage in LLM-generated tutoring feedback before deployment. Using 1,100 mathematics problems from the MATH dataset and 4,400 feedback items generated under Baseline, Direct, Socratic, and Encouragement styles, we compare rule-based, embedding-based, NLI-based, and two LLM-as-a Judge evaluators from different model families (Qwen2.5-32B and Gemma3-27B), and validate them against a 314-item human annotated gold standard. The results show that pedagogical style changes how answer-leakage signals appear in feedback and how reliably automated evaluators detect them. Rule-based and LLM Judge evaluators that perform well under Baseline, Direct, and Encouragement show consistent degradation under Socratic feed back. The primary LLM-Judge (Qwen2.5-32B) achieves overall F1 = 0.726 but drops to 0.427 on Socratic items, and a second LLM-Judge (Gemma3-27B) reproduces this degradation with Socratic F1 = 0.342. The four factorial prompt-sensitivity variants reach overall F1 up to 0.844, but their Socratic F1 remains at or below 0.390. These findings show that automated tutoring feedback auditing must be style-specific, prompt-calibrated, and human-grounded before deployment.
Foundational mathematics assessments can trigger anxiety and acute stress that discourage persistence in quantitative pathways. We introduce a video-story item format that places skill-matched mathematical problems inside short, relatable narrative videos. This video format represents a total-format design, combining relatable narrative, enacted video delivery, and temporal presentation; therefore, its observed effects reflect this combined design rather than any individual component in isolation. We evaluate whether this redesigned test experience is associated with lower physiological stress during problem solving. In a controlled study of lower-division undergraduates ($N=50$), participants completed a 36-item instrument designed around 12 skill-matched triplets—Video ($V$), Word ($W$), and Abstract ($A$)—while we continuously recorded multimodal affective signals during problem solving. We operationalize acute test stress as elevated autonomic arousal. We therefore analyze baseline-normalized cardiac dynamics (HR and IBI/RR-derived timing) and facial perspiration, and use facial expressions to characterize displayed valence. Cross-classified mixed-effects models accounting for participant differences and item difficulty show convergent reductions in physiological arousal across cardiac and facial-perspiration channels for the video-story format relative to conventional formats ($A$, $W$). Separate valence models show a shift toward more positive displayed valence. These affective advantages remain after controlling for time-on-item and occur without measurable loss in overall test performance. Crucially, we provide a practical procedure to derive participant-relative stress labels directly from baseline-normalized multimodal physiology, avoiding per-item self-report and enabling scalable dataset construction. Data, code, and materials are openly shared.
Engagement and persistence drive learning in online environments. Past research often proxied persistence with single, coarse engagement measures, such as time on task. However, recent research indicated more nuanced learner behaviors associated with productive and unproductive struggle. Here, we propose a longitudinal view based on log data in which engagement and persistence are expressed as repeated choices to start, continue, quit, or retry. Using two years of logs from bettermarks, an online environment for learning mathematics (N=2,129 students; >780,000 problem-set records), we estimated student propensities using Rasch-style models that adjust for content difficulty. By applying cross-validated factor analysis on propensities, reflecting decisions to start, continue, quit, and retry while working on problem sets, we identified three factors to predict future learning behavior. Completion Propensity predicts retrying and the selection of harder content despite lower accuracy; Error-Free Completion predicts higher first-attempt accuracy alongside a preference for easier content; and Activity predicts retry behavior while being associated with lower accuracy. The three-factor model explained substantially more variance than a unidimensional model in predicting future first-attempt accuracy (25% vs. 3%, respectively) and student-chosen content easiness (18% vs. 6%, respectively), though not retry behavior after failure. We discuss implications for effort-aware adaptivity and dashboards, contributing log-based indicators that can be computed for numerous online learning environments.
Collaborative problem-solving (CPS) activities encourage learners to share knowledge and complete complex tasks together. However, learners with insufficient prior knowledge may struggle to contribute effectively to these tasks. Appropriate support is therefore needed to help learners address challenges that arise during CPS. This study developed an AI personal assistant (AIPA) embedded with a ChatGPT-based large language model (LLM) to serve as a personal consultant in CPS activities by providing personalized knowledge, resources, and guidance. This design raises an important question about how AI assistance is associated with learners' interaction patterns and the quality of collaboration. By analyzing student dialogues and interactions with the AIPA, this study found no significant differences in CPS performance among high-, mixed-, and low-knowledge groups after using the AIPA. However, both mixed- and low-knowledge groups interacted with the AIPA more frequently than the high-knowledge groups. Analysis of question types further showed significant differences, with explanation-oriented questions being more common in low-knowledge groups. In addition, high-knowledge groups interacted minimally with the AIPA, low-knowledge groups engaged consistently, and mix- knowledge groups showed moderate interaction, with shifting question types across the CPS process. The results also showed that groups with lower prior knowledge relied more on information provided by the AIPA in their discussions. These findings suggest that the AIPA can serve as a supportive resource in CPS contexts and contribute to understanding students' interaction patterns with AI support in educational settings.
Personalized learning is nowadays frequently associated with the use of artificial intelligence (AI). This paper presents a systematic literature review focused on personalized learning supported by AI-based technologies, discussed through the lenses of the Activity Theory. The dataset combines manually selected studies until December 2024 with a supplementary set extending to October 2025. The paper explores the dialectic relationship between the subject and the object of an activity mediated by artefacts. The purpose of the research is to analyze current trends, technologies, theoretical approaches, and applications of AI in personalized learning. The review is focused on methodologically sound studies that use AI methods and algorithms to support learners, including evaluation procedures with concrete end-users and reasonable sample sizes. The discussion addresses the key aims of personalized learning supported by AI, the paradigmatic shift in personalization with the rapid development of AI, its role in ensuring flexible learning paths, and the main areas of application. It also covers challenges related to data quality, pedagogical soundness, and technology user interaction. This technological transformation is shifting traditional educational approaches towards more adaptive and learner-driven models.
Social robots are being used in education to aid both students and instructors across a wide range of disciplines and student cohorts. In this paper, we present a pilot human-robot interaction (HRI) study with a social robot using a large language model to autonomously present a lecture to a group of undergraduate university students. We developed the LectureGPT architecture, which consists of the novel integration of GPT-4 and EmoPrompt for a social robot to interactively deliver lecture content using both verbal and nonverbal communication. We investigated the effectiveness of LectureGPT in generating and presenting lecture materials directly from instructor notes using varying robot expressions. Results showed that LectureGPT can facilitate informative lectures and maintain student engagement. This research provides insights into the future use of robotic assistance for education in order to leverage them as interactive educational supports for non-expert users and expose students to state-of-the-art technology.
This study developed a dialogue-based Math intelligent tutoring system. Two tutoring modules—single-agent and dual-agent—were implemented based on the AutoTutor framework to support seventh-grade students in learning two-step fraction arithmetic through word problems. In the single-agent module, the agent plays the role of a teacher, whereas in the dual agent module, an additional agent plays the role of a peer student who interacts with the learner. To support adaptive instruction, a block-based matching method was applied to analyze students' responses, allowing the system to deliver feedback tailored to each student's performance. Diagnostic pedagogies include provoking cognitive conflict, problem simplification, representational teaching, and vicarious learning were incorporated into the instructional design. The developed system can be used for remedial instruction. A quasi-experimental study was conducted to compare the learning effectiveness among single-agent tutoring, dual-agent tutoring, and traditional teacher-led remediation. The experimental results indicated that dual-agent tutoring significantly outperformed the single-agent tutoring. These preliminary findings suggest that, within the current quasi experimental context, a dual-agent design incorporating peer interactions shows potential for enhancing learning effectiveness compared to a single-agent approach.
Generative Artificial Intelligence (GenAI) is profoundly transforming the research field, with many researchers embracing its potential to enhance their work. However, the integration of GenAI into research practices still presents non-negligible challenges. To address these critical issues and better understand the current state of GenAI use in research, we conducted a systematic review focusing on GenAI-assisted research in the social sciences area and screened 8,831 relevant articles, of which 126 were retained for in-depth analysis. Our findings reveal that GenAI has been widely adopted in empirical studies across various social science domains, demonstrating its value in supporting research processes, such as conceptualization and research design, data collection and analysis, and writing and editing. Drawing on insights from existing practices, we developed a guideline for the academic community to support the development of a broader consensus on GenAI-assisted research practices, particularly in the field of education. This guideline aims to better regulate and leverage the potential of GenAI in research, ensuring its ethical and effective application.
Driving skill acquisition remains a significant barrier to independence for many autistic adolescents and young adults. Although driving simulators provide a safe environment for practice, most systems function primarily as stand-alone tools and evaluate performance using discrete error counts or shortterm laboratory studies. This paper presents a telemetry-enabled driving training framework that integrates immersive simulation, multimodal behavioral sensing, and facilitator-guided instruction. The platform captures continuous vehicle control inputs and gaze-based attention signals during simulation and transforms these data into behavioral indicators describing control stability, attentional allocation, and safety-relevant events. The system was deployed as part of a multi-year driving training program delivered across multiple community-based service settings serving a broader population of neurodivergent learners; however, the present analysis focuses specifically on 47 autistic participants who completed pre- and post-program simulator assessments. Pre-post comparisons showed reductions in speed variability, steering oscillation, and non-gaze safety errors, indicating measurable simulator-based behavioral change within the driving simulator assessment context. These findings illustrate how telemetry-enabled simulation combined with facilitator-guided feedback can provide measurable insight into behavioral change during structured training. The system architecture also establishes a practical foundation for future adaptive and agentic AI capabilities grounded in educatorsupervised training contexts.
Knowledge Tracing (KT) dynamically monitors students' evolving knowledge mastery states through historical interaction sequences and predicts future performance. Existing KT methods often assume homogeneous temporal effects and therefore fail to fully capture the heterogeneous impact of multi-scale intervals on knowledge dependency evolution. To tackle this challenge, we propose the Multi-scale Temporal Graph Knowledge Tracing (MTGKT) model. First, we design a multi-scale temporal encoder that decomposes intervals into micro, meso, and macro scales and fuses them through soft assignment. Second, we introduce an adaptive forgetting mechanism based on the Ebbinghaus forgetting curve, with both knowledge-specific and student-specific parameters to model individualized memory decay. Third, we devise an enhanced dynamic graph neural network with structured neighbor sampling, skill co-occurrence features, and message passing, together with GRU and multi-head attention for temporal dynamics. In experiments, we conducted comparative evaluations between MTGKT and nine representative knowledge tracing models on four public datasets, and the results show competitive performance under the reported experimental settings.
Applying project-based learningfor teaching control systems often involves significant complexities related to hardware costs, safety, and preparation. SimuNEX is an open-source and interoperable simulation framework designed to facilitate the teaching and learning of controls engineering in a safe, engaging, and cost-effective manner. SimuNEX provides a 3-D environment that enhances students' understanding of spatial dynamics and system interactions, often not apparent from current signal and scope-based visual tools. This allows for a quicker and more intuitive debugging and iterative development of control algorithms. The platform supports realistic scenario-based learning and offers immediate visual feedback. This article discusses the design and implementation of SimuNEX, along with two application examples in an educational course on the modern control theory.
Rapid changes in the labour market have made reskilling and upskilling more relevant and important for career development than ever before. As the number of ways to obtain new skills and competences increases, so does the need for individual learning paths and support. As one of the most representative motivation and persuasive information systems, gamification has been employed as one potential approach for self-growth and skill development. This paper examines how different learners, i.e. students in higher education and lifelong learners who develop their skills and competences in work setting, value different gamified support mechanisms for individual competence development. For the study, a novel framework was developed that separates gamified support for competence development into two dimensions of temporality and openness. Utilizing the framework, a within-subject vignette based survey study was conducted with two learner groups (Higher education N = 544, Lifelong learner N = 285). Statistical analysis revealed that learners in higher education perceived overall value, i.e., usefulness, enjoyment and playfulness, in long term gamified support that allows flexibility and freedom to explore new learning opportunities. Lifelong learners perceived value in similar guidance but perceived additional utility in shorter-term gamified support that either guides them towards specific skills the learner should focus on or allows a flexible exploration and acquisition of skills at the learner's own pace.
Effective collaboration is essential for learning, especially when performing tasks that demand logical reasoning and problem-solving. Understanding how knowledge is co-constructed between individuals can enable valuable insights into the underlying cognitive processes. As teamwork increasingly shifts to online environments, there is a pressing need to investigate the dynamics of participant interaction beyond observable behaviors in depth. This study used a hyperscanning electroencephalography (EEG) approach, one of the most commonly used methods in social interaction research, to investigate interbrain synchrony (IBS) between dyads performing high school-level mathematics and computational thinking tasks of varying difficulty in a setting designed to simulate remote communication. IBS was predominantly observed in the frontal, temporal, and parietal regions and varied significantly across most frequency bands with respect to both task difficulty and completion time. Interestingly, an inverted U-shaped relationship emerged between IBS and task difficulty, with the highest synchrony observed during tasks of moderate complexity, suggesting that a balanced cognitive challenge may be optimal for promoting neural alignment. For the most demanding tasks, however, IBS was positively associated with longer completion times, implying that extended engagement with high-difficulty tasks may encourage sustained collaborative effort and neural coupling. These findings highlight IBS as a potential neural marker of effective joint problem-solving and shared cognitive engagement during collaborative tasks. This study has implications for educational design, team-based learning, and the development of neuroadaptive tools to enhance remote collaboration.
With the increasing prevalence of e-learning, accurately detecting moment-to-moment learner attention has become increasingly important for adaptive and personalized instruction. This study proposes an eye-tracking approach for evaluating attention in online video learning environments that complements conventional Areas of Interest (AOIs) with entropy- and velocity-based metrics that can be computed over the full screen. Nineteen university students watched diverse educational videos while their gaze data were recorded using a Tobii Eye Tracker 5 under focused and experimentally distracted conditions. Distraction was induced through a cognitively demanding serial-subtraction task, providing a more objective manipulation than self-reported mind wandering alone. Results showed that AOI-based measures capturing gaze shifts outside relevant regions, especially saccade out distance ratio and out saccade count ratio, significantly differentiated focused and distracted states. Non-AOI metrics further showed that distracted learners exhibited shorter fixations, more abrupt gaze-velocity peaks, and greater temporal irregularity, whereas attentive learners showed longer fixations, broader but more structured spatial exploration, and more consistent information sampling. To support interpretation under the modest sample, standardized effect sizes are reported alongside significance tests, and the findings are framed as exploratory validation. Overall, the results suggest that attention in e-learning is better characterized as stable, intentional visual exploration than as a simple reduction in eye movement. The proposed metrics provide a practical basis for more automated attention assessment beyond manual AOI setup. Future work should test generalizability across broader learner populations, content types, and multimodal sensing conditions.
