ECTEL 2026: Posters

1. BioStream: An open-source solution for collecting biosignals in educational environments through smartwatches

Mariano Albaladejo-González, Guillermo Vidal-Pina, Manuel
J. Gomez, Félix Gómez Mármol and José A. Ruipérez-Valiente

Abstract: Despite the widespread adoption of physiological sensors in user-oriented devices, researchers and educators cannot utilize them due to the lack of a user-friendly application for data collection and transmission. To address this problem, we introduce BioStream, an integrated, open-source infrastructure designed to streamline the acquisition of physiological signals using intuitive and affordable devices. This system comprises two independent components: SmartBioStream and ServerBioStream. SmartBioStream is a Wear OS application that enables the collection of physiological data through commercial smartwatches. Additionally, we provide a web platform, ServerBioStream, to receive, store, and download the data transmitted by smartwatches. The combination of both tools enables conducting educational research that requires the analysis of physiological signals. In addition to supporting research, SmartBioStream can transmit physiological data to other platforms, including learning management systems and educational simulators. This functionality enables monitoring of students’ physiological signals and the identification of potential risks, such as elevated heart rate. As an illustrative example, we have used SmartBioStream to send heart rate data to a Cyber Range, enabling the analysis of stress management skills during cybersecurity training. The flexibility, ease of use, and cost-effectiveness of both applications make them valuable tools for democratizing physiological data collection.

2. Reinforcement Learning in Serious Games: An exploratory analysis

Mariano Albaladejo-González, Manuel J. Gomez and José A.
Ruipérez-Valiente

Abstract: In recent years, Reinforcement Learning (RL) has emerged as a powerful approach for solving complex decision-making problems. RL enables the development of intelligent agents that interact with an environment by making sequential decisions. Rather than depending on labeled datasets, these agents learn directly from the outcomes of their decisions. For this reason, RL has promising applications in Serious Games (SGs) as they acquire knowledge directly through interactions with the game environment. This paper explores the application of RL in SGs, providing practical insights on how to incorporate such agents. First, we reviewed existing tools for integrating RL into SGs environments. Next, we analyzed how the format in which the agent receives information from the environment influences its training. Specifically, we assessed vector-based and visual observations across three environments of increasing complexity. The results showed a consistent advantage for vector-based observations in terms of convergence speed, training stability, and overall performance. Nevertheless, vector-based observations required a more extensive design before their integration, as they are highly dependent on the application scenario. Finally, we explored the applications of RL agents to enhance SGs. This work aims to encourage researchers and developers to integrate RL agents into SGs.

3. CAIMIL: A Multi-Agent Approach to Supporting Interdisciplinary Learning at Scale

Weiwei Li and Xiaojian Wang

Abstract: Interdisciplinary programmes are increasingly promoted in higher education to address complex societal challenges, yet their implementation is often constrained by fragmented curricula and the difficulty of facilitating scalable, re-al-time dialogue across student groups. While existing AI tools can support on-demand interaction, they remain limited in orchestrating structured, multi-perspective learning processes. This study presents CAIMIL (Collaborative AI Multi-Agent for Interdisciplinary Learning), a human-centered multi-agent system designed to support interdisciplinary dialogue through role-based agents and teacher-guided interaction. The system was co-designed with an interdisciplinary teaching team and deployed in a second-year programme at a Spanish university, involving approximately 200 students from primary education, psychology, social education, and social work. As a one-time in-class activity, CAIMIL complemented ongoing disciplinary learning around a shared case on inclusion. Both students and teachers perceived the system as usable, supportive, and well aligned with the course design. Classroom observations showed that students actively responded to agent prompts and engaged with multiple perspectives. However, some student’s contributions tended to be shorter and more reactive compared to AI-generated responses, suggesting a gap between participation and deeper interdisciplinary reasoning. These findings highlight the need for more adaptive scaffolding to support student agency and meaningful human–AI collaboration in technology-enhanced learning environments.

4. Plantagotchi: Learning with Living Systems through a Gamified Digital Twin

Alejandro Prats-Puerta, Bernardo Tabuenca, Vicente
García-Alcántara, Carlos Gilarranz-Casado and Alberto Cruz-Ruiz

Abstract: “This article presents a gamified digital twin-based learning ecosystem that enables learners to engage with real plant data through an interactive avatar. By coupling a hydroponic cultivation system with an Internet of Things platform and a mobile interface, the ecosystem creates a learning environment where environmental processes become observable and actionable. The pedagogical design follows an inquiry-based approach, supporting learners in exploring real-world problems, interpreting data, and making decisions based on system feedback. Preliminary implementations in agricultural and computer engineering contexts indicate that this approach can enhance engagement and foster systems thinking, particularly by making the consequences of learners’ actions visible over time. This work contributes to TEL research by (i) proposing a digital twin-based learning ecosystem for sustainability-oriented education, (ii) integrating IoT data and gamified interaction within an inquiry-based pedagogical design, and (iii) providing initial insights from its deployment across interdisciplinary learning activities. In doing so, the paper responds to current calls for more intentional, meaningful, and ecologically grounded learning technologies.”

5. A Learning Analytics Dashboard for Competency-Based Education to Support Class Council Meetings: A Design Study

Jonas Van Hove, Grzegorz Meller, Robin De Croon, Shahenaz
Najjar and Katrien Verbert

Abstract: Competency-based education (CBE), which emphasizes the development and assessment of student competencies rather than traditional grading, is increasingly adopted in anonymous secondary education. While this shift enables more holistic and continuous evaluation of student learning, it also introduces challenges for teachers in interpreting the resulting complex and heterogeneous data for decision-making, particularly during class council meetings. To address this, we present CLAIRE, a learning analytics dashboard designed specifically to support teachers in class council meetings. The dashboard integrates competency frameworks and student data into multi-level visualizations. The design was informed through a co-design session with 17 secondary school teachers, resulting in a system aligned with real-world requirements. We further conducted a qualitative user study with 5 teachers in an authentic school context. Findings indicate that the dashboard provides a clear overview of student progress, supports comparison between individual and class performance, and improves upon current reporting practices in class council meetings.

6. From Ambiguity to Agreement: Scaffolding Responsible GenAI Use in Doctoral Research

Patricia Khonje and Patricia Santos

Abstract: The rapid adoption of Generative Artificial Intelligence (GenAI) is transforming technology-enhanced doctoral learning while creating ambiguity around responsible use in research practices where disciplinary norms and ethical judgement are central. Institutional policies emphasize integrity but often lack contextual grounding in everyday human-AI research workflows, leaving doctoral researchers to interpret acceptable use independently. Existing research has focused largely on undergraduate contexts, over-looking doctoral education’s distinctive learning ecologies. This mixed-methods study examined doctoral researchers’ perceptions, practices, and regulation of GenAI use. Survey data from 152 PhD students across disciplines at a European university indicated near-universal adoption (95%), with frequent use supporting writing, coding, language mediation, and literature synthesis. While participants reported productivity and cognitive support benefits, they also expressed concerns about inaccuracy, overdependence, and ethical risks, and described routine verification practices. Qualitative findings show that GenAI is used as a cognitive support tool within hybrid human-AI research workflows rather than as a substitute for scholarly judgement. Findings reveal a practice–policy disconnect, doctoral researchers operate within fragmented supervisory learning environments where disciplinary conventions shape legitimacy and institutional guidance remains unclear, leading to individualized ethical decision-making and inconsistent disclosure. Grounded in these findings, we propose a participatory learning design approach to develop Doctoral Learning Agreements as a conceptual design artifact to scaffold supervisory dialogue on responsible GenAI use. This study contributes to Technology-Enhanced Learning research by showing how participatory learning design can support ethical sense-making and responsible human-centred GenAI integration in doctoral learning contexts offering a design direction aligned with intentional, mindful technology-enhanced learning.

7. From Black Box to Glass Box: GenAI in Non-STEM computing education

Lorena Gutiérrez-Madroñal, Juan Antonio Caballero-Hernández, Rubén Baena Perez and Manuel Palomo Duarte

8. Integrating Computational Thinking through Technology‑Enhanced Learning in Special Education: An Analysis of Self-perceived Teaching Competence and Contextual Barriers

Natalia Galbán Ojer, María Zapata Cáceres, Estefanía Martín Barroso and Marcos Román González

Abstract: This study analyzes self-perceived teaching competence in Computational Thinking (CT) in special education classrooms and the conditions that influence its pedagogical integration through Technology-Enhanced Learning (TEL). A quantitative, non-experimental, cross-sectional design was used with 33 special education teachers. A validated self-questionnaire assessed five dimensions: conceptual, pedagogical, technological-digital, ethical, and contextual. Descriptive analyses reveal an average level of self-perceived teaching competence. The ethical dimension stands out as having the highest score, indicating a strong awareness of the right to quality digital education that reduces inequalities. The lowest score was in the technological-digital dimension, highlighting shortcomings in the integration of CT using programming environments. The conceptual and pedagogical dimensions were at an average level, pointing to shortcomings in the curricular integration and assessment of CT. The contextual findings indicate significant institutional barriers, particularly the lack of practical guidelines and specialized training, although collaboration among teachers emerges as a facilitating factor. These findings are consistent with previous literature describing the gap between the recognition of CT as a cross-curricular competency and its effective implementation in special education classrooms. The study highlights the need for specific professional development, clear curriculum guidance, and organizational support to promote accessible and sustainable TEL practices. Strengthening these areas could improve equitable learning opportunities in CT for students with special educational needs.

9. Supporting Teacher Reflection and Instructional Insight through Multimodal Learning Analytics: An Exploratory Classroom Study with Reflecto

Kexin Yang, Yiliu Pan, Conrad Borchers, Nikol Rummel and Vincent Aleven

Abstract: Although substantial research has investigated how analytics-based tools can support teachers, an important question remains: how and what kinds of learning analytics can support teacher reflection, a fundamental mechanism for professional development. To address this question, we co-designed with teachers and developed Reflecto, a teacher-facing tool that supports hypothesis-driven reflection and exploration of multimodal learning analytics (MMLA). Reflecto synthesizes Intelligent Tutoring System (ITS) logs with teacher-position data to represent the interplay between student learning and teacher attention. We conducted an exploratory field study in six middle school mathematics classes with two teachers over two weeks to explore the reflection processes, the intended and actual changes in classroom practices, supported by Reflecto. By coding and analyzing transcripts of teachers’ reflection sessions, we gained an understanding of teachers’ reflection behaviors and their interactions with the tool. We found that Reflecto not only confirms and validates, but also extends educators’ knowledge and insights regarding their students and their teaching in the classroom. Reflection sometimes led to intentions to improve classroom practices. For one teacher, Reflecto contributed to student-specific instructional changes in the classroom, while the evidence remains limited for the other. Our findings suggest potential for a tool-supported process of hypothesis-driven exploration of MMLA to inform and challenge teachers’ understanding of their own classroom practice, and to change and improve teachers’ practice. This study contributes to a growing understanding of how MMLA data can meaningfully inform teachers’ instructional reflection and practice.

10. Is three a crowd?: The impact of Generative AI use on the teacher–student relationship in academic writing

Jacqueline Wong, Inge Baars and Anouschka van Leeuwen

Abstract: Generative AI (GenAI) has rapidly become embedded in students’ and teachers’ everyday practice, with academic writing emerging as one area of particularly notable impact. Students are increasingly using GenAI to support multiple phases of the academic writing process, including brain-storming, drafting, and editing. Likewise, teachers may use GenAI to assist them in the feedback they provide to students. While there is a growing body of research suggesting that students favor teacher-crafted feedback and view AI-generated feedback as of lower quality, considerably less is known about how the use of GenAI in teachers’ feedback practices on students’ academic writing influences the psychological need for relatedness among teachers and students. Additionally, while GenAI may offer teachers the benefit of reducing the time and effort when providing feedback, it is unclear whether doing so undermines the connection teachers might have with their students and vice versa, which is an important aspect of student motivation and learning. Therefore, in this study we examined several scenarios of GenAI use and compared students’ and teachers’ perceptions of relatedness satisfaction. Results showed significant differences in perceived relatedness satisfaction between students and teachers in conditions where teachers used GenAI to assist in providing feedback. Findings of the study contribute to a deeper understanding of the impact of GenAI from a self-determination theory perspective and provide insights into the potential benefits and cost of GenAI to teacher-student relationships.

11. Analytics for Agile Education: A Human-Centred Design Dashboard for Lecturers

Clara Gándara-González, María Jesús Rodríguez-Triana, Miguel A. Martínez-Prieto and Alejandra Martínez-Monés

Abstract: Agile education adapts professional agile practices to educational contexts to promote high-quality learning. It encourages students to progress toward learning outcomes through a sustained pace of activities and deliverables distributed over time. While this approach fosters continuous engagement, students may feel overwhelmed by the constant rhythm of work throughout the course and may struggle to distribute their workload evenly. In these cases, lecturers may face difficulties in understanding the students’ actual load and how they spend their time.
Curriculum Analytics (CA) could help lecturers to address this problem. In this paper, we present anonymized tool, a CA dashboard designed to support lecturers’ understanding of the workload and to inform the redesign of courses implementing agile practices. Anonymized tool enables the analysis of planning effort, timing, and workload associated with all learning activities in courses organised using the agile methodology Test-Driven Learning (TDL). We carried out an initial evaluation with four lecturers in a university course at the end of the 2024–25 academic year. Results indicate that anonymized tool helped them identify overlapped or poorly distributed activities, as well as some activities that led students to exceed the originally planned effort without a clear contribution to learning outcomes. Lecturers also reported using these insights to reflect on and improve the course design.

12. Structuring Critical Engagement in AI-Supported Learning: An Exploratory Multi-Aspectual Case Study

Sina Joneidy

Abstract: Recent pedagogical research examining students’ interaction with generative AI frequently assumes that learners possess the critical thinking skills required to evaluate AI-generated outputs. While many educational interventions encourage students to critique AI responses through fact-checking, rubric-guided assessment, or disciplinary reasoning practices, critical thinking itself is rarely explicitly conceptualised or systematically scaffolded. As a result, the enactment of critique often remains uneven across learners, depending on prior disciplinary knowledge, verification skills, and prompting strategies. This fragmentation suggests that current approaches promote questioning of AI outputs but lack a coherent mediating structure capable of stabilising critical thinking practices in AI-supported learning. More fundamentally, it indicates that critical thinking cannot be effectively cultivated through the direct application of reasoning frameworks alone, but requires carefully designed catalytic conditions through which criticality becomes observable in learners’ actions, judgments, and outputs.
Addressing this gap, the paper rejects the reduction of critical thinking to a mere technique and advances a conception of criticality grounded in whole-person knowing and situated, lived experience. Drawing on Dooyeweerd’s transcendental–systematic philosophy, criticality is understood as both fluid and structured: a mode of being that is not only existential, embodied, and action-oriented, but also structurally conditioned and normatively directed. In this view, critical being is expanded into structured critical being, whose thought, action, and experience are made meaningful within a multi-aspectual order of reality. This provides an ontological grounding in which reasoning and logic are situated alongside other aspects of everyday life, enabling criticality to function as a fully integrated mode of human knowing and acting.
Empirically, the study investigates whether multi-aspectual critique, grounded in Dooyeweerd’s Theory of Modal Aspects, can operate as such a mediating structure in generative AI–supported design tasks. The study was conducted as a qualitative, theory-informed case study of AI-supported learning within a small-cohort MSc Digital Marketing module, where students addressed a digital innovation challenge using the Design Thinking process. Students alternated between AI interaction and “unplugged” critical analysis phases, evaluating AI outputs through Dooyeweerd’s fifteen aspects before generating new prompts. Data consisted of participant–AI interaction transcripts and students’ off-AI analytical artefacts. Using directed qualitative content analysis, participant utterances were coded using Paul and Elder’s Critical Thinking Wheel and associated intellectual standards. Findings indicate that students who actively applied the aspectual framework demonstrated significantly richer patterns of critical engagement, including broader conceptual coverage, more frequent assumption surfacing, and stronger implication tracing. In contrast, instances where the aspectual analysis was either delegated to the AI or applied in an ad hoc manner showed substantially weaker evidence of critical reasoning.
The study suggests that multi-aspectual critique can function as a stabilising cognitive scaffold in generative AI–supported learning environments, shifting AI from a source of answers to an object of systematic interrogation, while grounding criticality in a structurally coherent and normatively guided framework. Rather than focusing on scale, it demonstrates how multi-aspectual critique operates as a structured catalyst that renders criticality visible in learners’ reasoning, decisions, and outputs, linking philosophical grounding, pedagogical design, and observable behaviour in AI-supported learning.

13. Who benefits from AI in TEL? Rethinking AI Literacy through Capital Theory

Alexandra Gössl and Robert Pham Xuan

Abstract: Recent discussions in TEL highlight the potential of generative AI tools to enhance accessibility, provide adaptive tutoring and enable personalized learning pathways. At the same time, emerging digital divide debates acknowledge uneven access conditions, differential AI literacy, and concerns that AI use may reproduce existing inequalities. These tensions point to an underlying issue regarding the distribution of educational benefits. This, in turn, raises a more fundamental question: who is actually able to translate these potential advantages into meaningful educational outcomes?
This paper argues that benefiting from generative AI in TEL requires more than access or basic operational skill. It involves the ability to convert online activities into beneficial offline outcomes. Drawing on sociological theory — particularly Bourdieu’s concept of capital as resources that are distributed unequally and can be converted into social advantage — the presentation will conceptualise AI literacy as a form of digital capital. Therefore, instead of viewing AI literacy purely as a technical skill, as is often the case in competency-based frameworks, we suggest considering it a form of digital capital that encompasses strategic judgement, evaluative abilities, and a disposition towards reflective and critical engagement with AI systems.
From this perspective, AI-enhanced TEL environments may not generate uniform benefits, but instead their educational value is mediated by the distribution and convertibility of digital capital. Reframing AI literacy through capital theory therefore provides a conceptual lens for examining how advantages in AI-supported learning are socially structured rather than universally accessible. Such a perspective may help to identify the unequal conditions under which AI-supported learning actually takes place and can inform more inequality-aware approaches to TEL-design and policy discussions. In this sense, the paper contributes to ongoing debates about how learning technologies might be developed with intention, by drawing attention to the social preconditions that shape who is able to benefit from them.

14. Exploring the Role of Personality-Driven Large Language Models to Facilitate Collaborative Problem-Solving

Ishari Amarasinghe, Jacopo Amidei, Rubén Nieto Luna and Salvatore Romano

Abstract: Computer-Supported Collaborative Learning (CSCL) enables knowledge co-construction and problem-solving through digital technologies. While traditionally focused on human–human interaction, CSCL is increasingly expanding to include collaboration between humans and Artificial Intelligence (AI) agents. However, research that positions Large Language Models (LLMs) as active collaborators rather than passive tools remains limited. In addition, such AI agents are rarely personalised to match the specific characteristics of individual users. In this poster, we investigate the role of personalised AI in human–AI collaboration, focusing specifically on the impact of personalised LLM-based agents. Personalization was operationalized through personality traits. Twenty-four participants collaborated with either a personality-aligned, personality-inverted, or non-personalised LLM agents to solve a Sudoku puzzle. Preliminary findings suggest that personality-aligned agents supported collaboration compared to both personality-inverted and non-personalised conditions. Workload assessments indicate a trend toward higher mental demand, temporal demand, effort, and frustration when the agents were not personalised.

15. Supporting Epistemic Agency Through Multi-Theory AI Explanations: A Mindful Learning Analytics Dashboard

Wenting Sun

Abstract: As generative AI becomes embedded in learning analytics dashboards, these systems increasingly shape how educators interpret and act upon student data. Under the European Union Artificial Intelligence Act, such systems are classified as high-risk, raising critical questions about transparency, accountability, and human oversight. Existing dashboards often prioritise predictive outputs or single authoritative explanations, potentially reducing educators’ role to passive recipients of algorithmic judgments. Yet interpreting learning data is inherently theory-laden and context-dependent. This study proposes a mindful learning analytics dashboard prototype that supports epistemic agency through multi-theory AI-generated explanations. The prototype presents scenario-based cases and generates alternative interpretations grounded in Self-Regulated Learning (SRL) and the ICAP framework, enabling reflective comparison rather than automated acceptance. An exploratory qualitative study was conducted using semi-structured walkthrough interviews with educators and educational stakeholders. The analysis examines how participants evaluate, compare, and negotiate competing AI-generated interpretations. Findings show that participants maintain interpretive authority, calibrate trust through theory and evidence, and engage with AI as a reflective rather than decision-making partner. The study contributes a design approach that shifts AI from providing answers to supporting interpretation, aligning with the goals of mindful Technology-Enhanced Learning.

16. Listening Comprehension in Virtual Reality: An Eye-tracking Study on Contextual Cues

Ting-Yu Liu, Pauline Frick, Tobias Appel, Marc Halfmann, Kateryna Derkach, Peter Gerjets, Detmar Meurers and Andreas Lachner

Abstract: The descriptive study aimed to understand the perceptual processes under-lying auditory comprehension in VR and was built on a previous VR study we conducted, where the degree of contextual support was systematically varied. The previous study showed the effectiveness of visual context in VR towards listening learning in English as a foreign language and indicat-ed that richer contextual support by actors in a virtual environment further enhances listening comprehension. Leveraging the effectiveness study, the current study adopted its experimental design and explored learners’ atten-tional processes while engaging with contextual cues in VR via eye-tracking to comprehend how they perceive given cues in authentic envi-ronments. To observe learners’ eye movements, 24 context-related areas of interest (AOIs) were set in our virtual environments as well as three AOIs on virtual actors. The collected eye-tracking data reveal that the presence of virtual actors concentrates learner gaze and governs attention shifts between social and contextual stimuli. The findings tentatively suggest that social presence may influence cognitive engagement during immersive language learning, pointing toward a possible mechanistic account of how actor-mediated and contextual cues could support knowledge construction in VR-based educational settings.

17. Eliciting Ethical Reasoning for AI in Learning Analytics: A Scenario-Based Study Using Epistemic Agency

Wenting Sun

Abstract: The increasing integration of artificial intelligence (AI) into educational environments raises significant challenges for data governance, including issues of privacy, fairness, transparency, and accountability. While existing frameworks such as FATE provide normative guidance, less is known about how educators and educational stakeholders interpret and negotiate these ethical principles in practice. This study investigates how educators and stakeholders reason about AI governance in learning analytics through a scenario-based elicitation approach. We developed a lightweight research prototype that combines a literature-derived scenario knowledge base with a generative AI component to prompt reflective engagement with governance dilemmas. Using this system, educators and educational stakeholders participated in scenario-based interviews exploring data use, consent, algorithmic decision-making, and institutional responsibility. Findings show that participants do not rely on abstract ethical categories when reasoning about AI governance. Instead, their judgements are shaped by four key patterns: (1) purpose-based and context-dependent evaluation of data practices, (2) positioning of human judgement as a necessary boundary condition for AI use, (3) attributing governance responsibility primarily to educational institutions, and (4) identifying broader educational harms beyond privacy. These results suggest that ethical reasoning in educational AI emerges through situated negotiation of epistemic agency rather than through predefined governance frameworks. The study contributes a conceptualization of ethical reasoning as epistemic agency, a scenario-based methodological approach for eliciting such reasoning, and design implications for developing reflective, context-aware AI governance tools in education.

18. Digital Mediation: EFL Writers’ Reflections on Dynamic Assessment through Generative AI Tools

Tuba Özturan

Abstract: This qualitative study examines the reflections of EFL (English as a Foreign Language) writers on integrating Dynamic Assessment (DA) principles with Generative AI (GenAI) tools. While previous research highlights the pedagogical potential of GenAI in providing personalized feedback, little is known about how DA can be operationalized through such tools and how learners perceive this integration. In L2 writing, adaptive and individualized feedback is particularly valuable because learners exhibit diverse strengths and weaknesses. DA is based on Vygotsky’s Sociocultural Theory and highlights the importance of addressing learners’ unique differences in the learning process. In this regard, it offers a learning-oriented assessment framework that merges assessment with instruction in collaborative interaction between teachers and learners. However, its iterative and systematic nature often presents challenges for teachers and might warrant blending DA with technology (Davin & Herazo, 2020). GenAI tools may help address these challenges, yet research combining DA and GenAI in L2 writing, particularly examining the L2 writers’ perspectives, remains scarce.
This study investigates the experiences of 19 B1-level EFL learners enrolled in an L2 writing course. Over the semester, participants completed eight paragraph-to-essay-writing tasks across different genres. Following each independent writing task, they engaged with ChatGPT-4, which was guided to act as a DA provider using a structured prompt adapted from Author. ChatGPT-4 delivered personalized feedback on both global and local aspects of writing, tailored to individual learner performance and responsiveness. After each task, students submitted short reflections of their interaction with DA-based ChatGPT-4. Approximately 150 reflections will be thematically analyzed. The preliminary findings reveal both benefits and challenges in employing ChatGPT-4 as a DA provider, offering insights for L2 writing teachers and language testers seeking to integrate learning-oriented assessment with GenAI support.

19. Towards a Mapping for AI-Based Eye-Tracking Analytics in Multimedia Learning

Konstantinos Tsiakas, Francesca Zermiani, Deniz Iren, Nghia Duong-Trung, Michael Raschke, Roland Klemke, Milos Kravcik, Ladislao Salmerón, Halszka Jarodzka and Leen Catrysse

Abstract: Supporting students’ reading comprehension is demanding for teachers, as they have limited insight into the reading processes of each student while they are reading. Although eye-tracking and artificial intelligence (AI) approaches offer strong potential to generate such insights, they remain difficult to deploy and apply in actual classroom practice. A key limitation lies in the lack of a structured approach to translate gaze data into pedagogically sound, classroom-ready, actionable insights for teachers. In this paper, we address this gap with a proof-of-concept mapping that connects the Cognitive Theory of Multimedia Learning (CTML) processes—selecting, organizing, and integrating—and their associated eye-tracking metrics with specific AI methods and HCI-driven visualizations. In this ongoing work, we present an initial mapping derived from a literature review: we identify eye-tracking metrics associated with CTML processes and propose preliminary connections to AI methods and HCI-driven visualizations. Drawing on this mapping, we illustrate our approach with a concrete example from a PISA task that demonstrates how AI-driven interpretations, combined with HCI visualizations, can make such insights accessible in educational contexts. We argue that this structured approach is a necessary step toward designing more targeted applications—moving from raw gaze data to feedback that teachers can use. By grounding technical and design choices in multimedia learning science, this mapping aims to support the development of AI-enhanced tools that are both technically sound and pedagogically relevant.

20. Towards Learner Agency-Preserving Feedback in Generative AI: A Participatory Design Study

Yueling Fan, Jaqueline Wong, Olov Engwall and Olga Viberg

Abstract: The widespread use of commercial conversational generative AI tools in higher education has raised concerns about students’ overreliance on them and the potentially detrimental effects of cognitive offloading on learning. These concerns highlight the need to support students’ self-regulated learning (SRL) by fostering metacognitive processes that enable them to monitor and reflect on how these tools are used to aid rather than take over the learning process. Although feedback is a promising means to improve SRL, designing feedback that promotes reflection while preserving learner agency remains a challenge in AI-mediated learning. A central design challenge lies in the tension between students’ inclination for executive help-seeking and the need to design feedback that redirects learners toward agency and ownership of their learning process. This work-in-progress study adopts a participatory design approach to explore the design of human-centered feedback to support learner agency in programming- and academic writing tasks. Through a co-design workshop, 18 students used a card-based toolkit informed by the theoretical lens of agency and prior research on SRL and AI-supported feedback. Initial findings demonstrate key design dimensions for integrating SRL-supportive feedback into conversational AI interactions. These dimensions focus on fostering reflection, preserving learner control, and encouraging active engagement with feedback. As an initial contribution, we introduce a framework-informed design toolkit aimed at supporting the development of human-centered, agency-preserving feedback in human-AI learning practices.

21. From Eye Movements to Instructional Insights: Co-Designing an AI Dashboard with Teachers

Konstantinos Tsiakas, Leen Catrysse, Siem Buseyne, Tine van Daal, Vincent Donche, David Gijbels and Halszka Jarodzka

Abstract: Many important learning processes, such as reading, are not directly observable to teachers. Eye tracking can make these hidden processes visible, but the data produced is too complex for direct classroom use by teachers. Artificial Intelligence (AI) can interpret these complex data and convert them into practical, actionable insights that support teachers. This paper presents the exploratory co-design phase of an AI-assisted dashboard that analyzes student eye-tracking data and provides data-informed insights to teachers through visualizations of reading behaviors alongside AI-generated alerts and recommendations. During co-design sessions, teachers engaged in guided exploration of a prototype that simulates AI model outputs using artificial student data, including detection of skimming and perceived reading difficulty, clustering of students into reading profiles, and large language model (LLM) generation of teaching insights. A think-aloud protocol was used to gather feedback on the potential utility of the prototype. A preliminary deductive thematic analysis from sessions with 10 teachers reveals seven core teacher needs: actionable narratives, surface control, teacher support, deep agency, on-demand explainability, pedagogical integration, and teacher-envisioned functionalities. We discuss the implications of these findings for the subsequent phases of the development of a tool for classroom practice.

22. Embedding Self-Regulation in a Quiz-Based Online Learning Environment: A Design-Based Research Study

Amir Ansari, Suvechhaya Shrestha, Abi Atamanesh, Rana Coskun, Niri Gala, Adrijana Krebs, Sebastian Dennerlein, Wilko Rohlfs and Alexander Steinmaurer

Abstract: In online learning environments (OLE), structured guidance is crucial for quiz-based learning to avoid superficial activity where students attempt questions without reflecting on their learning progress and planning their next steps. Effective learning in such environments requires self-regulated learning (SRL) skills. However, most online quizzes do not prompt learners to set goals (forethought) and evaluate their understanding (self-reflection), which allows them to regulate their subsequent learning activities (e.g., selecting learning materials) in the OLE. This paper presents a Design-Based Research (DBR) study for developing and evaluating an SRL scaffolding module embedded in an OLE in the Heat and Mass Transfer course. This module structures quiz interaction around SRL phases by providing goal-setting prompts before the quiz and AI-generated reflection prompts afterwards and targeted access to relevant learning materials. The evaluation employs an explanatory sequential mixed-methods design. Quantitative analysis examines how students’ SRL skills relate to and predict their interaction patterns and perceived usefulness of the module, while qualitative accounts provided explanatory insights. Results show that self-evaluation is the strongest predictor of interaction behaviour, and overall SRL significantly predicts perceived usefulness. Qualitative findings indicate that reflection and resource access support learning effectively, whereas goal-setting prompts are often perceived as disruptive. These findings provide design-oriented insights for integrating SRL scaffolding into quiz-based OLE, highlighting the importance of reflection, strategic support, and context-sensitive implementation of planning features.

23. Mapping the Clinical Reasoning Process in Medical Simulations Using Multi-Cluster Multi-Layer Transition Network Analysis

Chengxin Zhang, Sonsoles López-Pernas, Wentao Wang and Mohammed Saqr

Abstract: Clinical reasoning (CR) is a foundational skill in medical education, where improving students’ CR skills remains a central objective for training programs. Achieving this goal requires an in-depth understanding of the dynamics of CR and how it unfolds in doctor-patient interactions. In this study, we investigate the dynamics of students’ CR processes as they emerge during simulated patient–doctor conversations. We analyze a dataset comprising 213 clinical simulation dialogues. To capture the sequential and relational structure of these interactions, we employ Multi-Cluster Multi-Layer (MCML) Networks, an extension of Transition Network Analysis (TNA) that accounts for multiple-layer node classes (in our case, distinguishing between dialogue intent and clinical goal). In doing so, we model how reasoning-related actions evolve throughout the medical simulation from the initial history-taking to the diagnosis and treatment. Our findings provide insight into the interactional patterns that characterize CR and demonstrate the value of MCML for examining complex, multi-layer processes in medical education.

24. Assessing Procedural Knowledge and Learner Experience in VR Safety Training Using Multimodal Learning Analytics

Timothé Bonhoure, Mélina Verger and Élise Lavoué

Abstract: Virtual reality (VR) environments offer strong potential for training in high-risk domains such as medical and chemical laboratory settings. Their effect on learning is usually measured using self-reported questionnaires. However, these measures do not adequately capture how learning evolves over time in immersive environments. By combining physiological signals, eye-tracking data, and behavioral traces, multimodal learning analytics (MMLA) provide a richer and more integrated understanding of learning processes. Beyond learning assessment, MMLA is also highly relevant for risk prevention and incident management, as it may capture key indicators such as learners’ stress and attention. In this paper, we present an experimental framework that synchronizes multimodal data collection in a virtual chemistry laboratory training environment. We investigate how the analysis of these data can inform on both the development of procedural knowledge for risk prevention and incident management, and the overall learner experience (flow, cognitive load, presence). This work ultimately aims to characterize the dynamics of learning indicators to support the adaptive design of training scenarios for safer practice.

25. Toward Enhancing Tutor Feedback Quality through AI-Assisted Training and Self-Reflection

Imen Azaiz, Sven Strickroth and Tatjana Riedmaier

Abstract: Personalized formative feedback is particularly crucial for novice learners in large introductory programming courses. However, tutors are often students themselves, and their written feedback tends to be brief, unstructured, or insufficiently motivating, thereby limiting its effectiveness. Prior work has explored AI-generated feedback for students, but less attention has been paid to supporting tutors in improving their own feedback practices. This poster presents an AI-assisted reflective training prototype and reports on an exploratory study investigating its potential to help tutors reflect on and improve their written feedback. The prototype combines a micro-tutorial on feedback categories with LLM-generated feedback that tutors receive on their own writtenfeedback to students’ programming submissions. The approach follows a two-step process in which the LLM first evaluates the tutor’s feedback (e.g., in terms of elaboration and motivation) and then provides targeted suggestions for improvement. A preliminary study (n = 13) shows that tutors who completed the tutorial wrote substantially more elaborated feedback after training, with elaboration rates nearly doubling from 46 % to 92 % and average word count increasing by 82 %, and tutors reported more positive attitudes toward the feedback process (SU S = 84). These findings suggest two main use cases: as a training component at the beginning of a course and as an on-demand reflective support tool within authentic feedback processes.

26. Auditing Bloom with Open-Weight LLM Councils for AI-Mediated Assessment

Alessio Ferrato, Carla Limongelli, Daniele Schicchi and Davide Taibi

Abstract: While question banks annotated with Bloom’s Revised Taxonomy are vital for Computer Science (CS) assessment design, assigning cognitive levels to programming tasks is inherently ambiguous. These interpretive ambiguities suggest that legacy resources may benefit from further validation and refinement. We demonstrate this flaw through an empirical audit of 102 legacy CS questions, revealing only marginal agreement among four domain experts (Fleiss’ kappa = 0.2497). To mitigate this classification noise, we propose an AI-mediated auditing tool utilizing a multi-agent council of open-weight Large Language Models. Experimental results show that this council architecture significantly outperforms individual models, reaching 52.9% accuracy against the expert consensus. The council’s voting variance serves as a diagnostic metric, empowering educators to flag and refine ambiguous items and improving the reliability of Technology Enhanced Learning environments.

27. Designing Support for Written Reflections using GenAI Summaries and Chatbots

Mikaela E. Milesi, Viktoria Pammer-Schindler, Vanessa Echeverria, Sonsoles López Pernas, Mohammed Saqr, Riordan Alfredo, Yueqiao Jin, Jie Xiang Fan, Dragan Gašević, Yi-Shan Tsai and Roberto Martinez-Maldonado

Abstract: Reflection is a critical process for learning as it allows students to assess their knowledge and skills while deepening their understanding of learning material. In healthcare education, simulations provide students with the opportunity to cultivate clinical and soft skills by reflecting on their decision-making and actions. Yet, reflection on evidence captured in these simulations is often scarce. Learning analytics (LA) dashboards can support reflection by visualising multimodal data. However, if students are unable to comprehend the data, they may be unable to derive actionable insights to improve their practice. It is thought that generative AI (GenAI) can address this challenge by providing simplified explanations of data. However, little is known about the extent to which GenAI interventions (such as using chatbots or AI-generated summaries) can lead to constructive reflections on LA data. We conducted a comparative study under authentic conditions with 40 nursing students who wrote reflections after a clinical simulation, supported by an LA dashboard augmented with either a chatbot or AI-generated summary. Transition network analysis on the student reflections indicates that those in the AI-summaries condition were more likely to centre their discussions around planning future actions, while reflections supported by chatbots engaged with a broader range of aspects, including insights from the simulation, future planning, and self-assessment.

28. On the Impact of Feedback Length on Reading Attention

Ana Tibau-Flores, Sergi Solera-Monforte, David Arnau, José Antonio González Calero and Miguel Arevalillo-Herráez

Abstract: Cognitive Load Theory posits that textual density impinges on the limited working memory of learners, a constraint that is particularly critical in primary education. When the cognitive cost of processing feedback outweighs its perceived value, students risk shifting from deep comprehension to superficial scanning. In this paper, we define and empirically investigate the Attention Ceiling: the critical threshold where additional text length yields diminishing instructional returns. Using the Hypergraph-based Intelligent Tutoring System, we deployed a granular interaction mechanism to analyze the real-time allocation of attention against feedback message length. Our results identify a distinct inflection point. While reading time initially shares a linear relationship with text length, it paradoxically decreases beyond a specific saturation threshold, signaling a behavioral transition to heuristic scanning. These findings provide actionable constraints for the design of Intelligent Tutoring Systems, advocating for strict length penalties to ensure feedback remains within the cognitive grasp of young learners.

29. Toward the Alignment of Understandability, Reasoning and Explainability in TEL

Sonja Klein and René Röpke

Abstract: Artificial intelligence (AI) enabled learning technologies (LT) are rapidly spreading in classrooms, yet their opacity makes it difficult for teachers and students to align them with pedagogical goals. Existing work on explainable AI (XAI) in education is often criticized for being model‑centric, emphasizing what can be explained rather than what stakeholders need to understand in situ. In line with the call for human-centered XAI in education, we propose to focus on the Alignment of Understandability, Reasoning, and Explainability in Technology‑enhanced Learning (AUREtel) – a triadic, relational model that treats mindful technology integration as the balanced interplay of three bidirectional explanation edges among LT, teachers, and students. We hypothesize that mindful LT integration depends jointly on the quality of explanations across all three relationships. By shifting from model‑centricity to explainability as a pedagogical affordance, treating explanations as designable pedagogical artifacts, AUREtel aims to guide human‑centered design and evaluation of explanations. With this conceptual work, we propose to further discuss the relations among stakeholder-specific understandability, pedagogical reasoning, and explainability.

30. The Educational Economics Simulation Hub Power Law Signatures in Student Learning, Engagement and Strategy

Marcello Silvestri

Abstract: This paper presents the design, implementation, and empirical evaluation of the Educational Economics Simulation Hub, a platform of 24 economic simulation games developed and deployed in secondary school classes as both a teaching tool and a research instrument, combining evidence-based learning design with data-driven behavioural analytics. Each simulator immerses students in a concrete economic role, provides immediate feedback, and feeds a nine-competency assessment framework monitored via a teacher dashboard. Analysing 12,349 strategic decisions from 57 students across four eligible simulators, we find preliminary evidence consistent with three Power Law distributions: the Power Law of Practice (β = 0.528, R² = 0.952), Pareto engagement (α = 1.839, R² = 0.846), and Zipf strategy frequency (α = 2.051, R² = 0.829). The platform was built through vibe coding—natural language collaboration with generative AI—showing that teachers without high programming skills can design research-grade experiential learning environments. These findings are exploratory (N = 57) and require larger-scale replication.

31. Educator-AI Collaboration for Facilitating System 2 Thinking: an LLM-powered Agentic Instructional Support System in Higher Education

Adrijana Krebs

Abstract: This research investigates how educators and AI can collaboratively support students’ System 2 thinking in higher education. Grounded in Dual Process Theory, the research focuses on the design, implementation, and evaluation of a multi-agent instructional support system that promotes deliberate, reflective, and analytical reasoning. The study responds to two related challenges: students often need adaptive scaffolding to engage in deeper reasoning, while educators face limited time, large student cohorts, and increasing pressure to integrate technology meaningfully into teaching. Using a mixed-methods and design-based research approach, the dissertation examines educators’, students’, and educational developers’ perspectives on the evolving role of educators, analyzes teaching strategies that foster System 2 thinking, and develops an AI-supported instructional system. The proposed system uses adaptive feedback, hints, and tutor-like guidance to support learners while extending educators’ capacity to provide individualized support. The research contributes to AI-enhanced instructional design by translating Dual Process Theory into a multi-agent learning environment and clarifying how educators’ pedagogical roles can be preserved and strengthened in technology-mediated education.

32. Teacher-Facing Multimodal and AI-Supported Feedback Orchestration for Oral Presentations

Alvaro Becerra

Abstract: Learning is inherently multimodal, and providing rich, personalized feedback requires capturing and interpreting diverse evidence about students’ learning processes and performance. In this context, Multimodal Learning Analytics (MMLA) shows strong potential for capturing complex aspects of learning by integrating biometric, contextual, behavioral, and interaction data. However, current MMLA approaches face challenges related to scalability, data integration, interpretability, and the limited translation of analytics into feedback that teachers can use in authentic classroom settings. This doctoral research, situated within the Technology-Enhanced Learning domain, explores the integration of collaborative human assessment, Machine Learning (ML), MMLA, and Generative Artificial Intelligence (GenAI) to support teacher-facing feedback orchestration. The empirical focus is oral presentations in higher education and to address this context, the dissertation proposes MOSAIC-F, a framework that supports the oral presentation lifecycle through five stages: (0) pre-presentation slide feedback, (1) collaborative rubric-based assessment to obtain human-based insights, (2) multimodal data collection and ML-based extraction of learning and performance indicators, (3) GenAI-supported feedback generation, and (4) self-assessment with feedback visualization. The expected contributions include scalable multimodal data collection methods, interpretable ML-derived indicators, and GenAI-based feedback mechanisms that help translate complex multimodal evidence into pedagogically meaningful feedback for teachers and students.

33. Design Principles for Hybrid Learning in Vocational Education: A Multi-Case Co-design Study

Gabriela Brezowar, Tobias Ley and Marlene Hopfgartner-Wagner

Abstract: Synchronous Hybrid Learning (SHL) enables co-located and remote learners to participate simultaneously in shared learning activities. While SHL has gained increasing attention in recent years, empirical evidence from voca- tional secondary education remains limited. This study investigates four co-de- signed SHL implementations conducted across Austrian vocational schools. The study employed a multi-case mixed-methods design combining quantitative, qualitative, and process-oriented data sources. Data were analysed through cross- case comparison and triangulation to identify recurring implementation patterns and derive transferable design principles. The findings revealed four recurring patterns across the analysed SHL implementations: (1) authenticity as a driver of engagement, (2) social presence as a challenge requiring deliberate design, (3)technology as both an enabler and a constraint, and (4) coordination as an im- portant success factor. Based on these findings, a set of design principles was derived addressing technological infrastructure, role allocation, moderation, learning phase design, external collaboration, and active participation.

34. Designing and Instructing AI-based agents for Learning Computational Thinking in Higher Education

Saba Soleimani

Abstract: Computational thinking (CT) has gained importance across a wide range of disciplines in higher education, yet many students, particularly those from non-computer science backgrounds, struggle to develop CT skills through conventional instruction. AI-based pedagogical agents offer new opportunities for adaptive and interactive learning support, but current approaches mainly position AI as a tutor and provide limited evidence on reciprocal human-AI learning. This PhD project investigates two complementary AI-based learning approaches in higher education: PIA, a Peer Instructional Agent acting as a tutor, and AILA, an AI Learner Agent taught by students. The project follows an Educational Design Research approach and is organized in two phases. In phase one, in two studies the two types of pedagogical agents are developed and evaluated. In phase two, an empirical investigation on learning with the two pedagogical approaches is carried out. It looks at learning from a holistic viewpoint and investigates the complex interplay between learner characteristics and preferences, learning outcomes like performance and satisfaction, and pedagogical approaches. The expected contribution is a theoretically grounded and empirically evaluated framework for designing reciprocal AI peer agents for CT learning in higher education.

35. Redefining scientific skills education: A Dutch national Delphi study on whether and when medical students should use Generative AI

Remco Jongkind, Birgit van Berlo, Sarah Otto, Florine van Driessen, Tommy Pattij and Suzanne Geerlings

Abstract: The rapid proliferation of Generative AI (GenAI) has disrupted higher education, challenging traditional written assessments and learning objectives. Since medical education must adhere to national guidelines and the 2020 Dutch National Blueprint for Medical Education predates these breakthroughs, educators lack national guidance on the incorporation of GenAI in scientific development learning goals and required GenAI literacy.
To establish actionable consensus, we conducted a Delphi study across all eight Dutch University Medical Centers involving curriculum coordinators, researchers, and assessment specialists. Through iterative surveys and meetings (>75% agreement threshold), the panel evaluated 32 scientific and 11 GenAI literacy goals. We investigated which objectives require GenAI-free versus GenAI-integrated approaches, the optimal timing for integration, and prerequisite skills.
Findings revealed a pedagogical bifurcation. Experts agreed foundational skills, such as academic attitude and experimental conduct, must be developed predominantly without GenAI. Conversely, scientific communication tasks can integrate GenAI earlier, provided students first demonstrate unmediated competence. To facilitate this, the panel validated 10 prerequisite GenAI literacy goals.
Consequently, we propose two operational approaches: 1. GenAI-Free Zones: Building cognitive foundations through synchronous, high-validity assessments like oral defenses (“”learned conversations””) and assignments like in class micro-writing. 2. GenAI-Integrated Learning: Preparing for the workplace by evaluating both final products and human-AI collaboration.
This approach protects critical reasoning while equipping professionals with digital competencies. Attendees will receive an evidence-based mapping of GenAI integration for learning goals, a consensus-backed set of 10 GenAI literacy learning goals, and a methodological blueprint for establishing internal consensus in their own educational programs.

36. COLLVIT: Investigating Collaborative and Competitive Feedback in AI-Supported Psychomotor Learning

Jon Echeverria

Abstract: This doctoral research investigates whether artificial intelligence (AI) may contribute to a better development of psychomotor skills through collaboration using the example of collaborative strategies for giving feedback. Most applications of AI in supporting the learning of psychomotor skills focus on the use of AI to correct the errors made by individuals. Many embodied activities however need coordinated activity, synchronised action and mutual adaptations from participants. This research is developed by combining different areas of knowledge in a technological architecture: Technology-Enhanced Learning (TEL), Computer-Supported Collaborative Learning (CSCL), and sociocultural learning perspectives. This technological architecture serves as a reference for the psychomotor analysis of two people practicing kihon kumite. The aim is to analyze the differences in results, motivation, and other key teaching parameters depending on the type of feedback used (individual, collaborative, competitive).

37. Measuring Multi-Skill Learning Progress with Optimal Transport

Diana Nurbakova

Abstract: Standard learning analytics metrics treat skills as independent dimensions, discarding the developmental relationships between them. We propose using the Wasserstein distance with pedagogically grounded ground metrics to measure multi-skill learning progress. The ground metric encodes a skill hierarchy, making the cost of reshaping a student’s profile depend on the structural distance between skills. Validated on 79K professional translation quality annotations and calibrated synthetic trajectories, the approach achieves a 5.7$\times$ improvement in learner archetype discrimination over Euclidean distance and surfaces trajectory distinctions invisible to pointwise and distribution-based alternatives alike.

38. xEST: a Modular Architecture for Designing and Enacting Ubiquitous Learning Situations Across Multiple Domains Based on Linked Open Data

Pablo García-Zarza, Guillermo Vega-Gorgojo, Eduardo Gómez-Sánchez, Miguel L. Bote-Lorenzo and Juan I. Asensio-Pérez

Abstract: The technology-enhanced learning community seeks to apply technological advances to improve educational processes. Among the different fields where these advances can be applied is ubiquitous learning, which enables the design and enactment of learning experiences that take place anytime and anywhere across spaces. This approach makes it possible to tailor learning designs to students’ contexts and needs, which, among other benefits, fosters their critical thinking. However, this level of resource customization increases teachers’ workload. This workload can be reduced if teachers are able to reuse existing resources. In this regard, semantic technologies emerge as an opportunity to enable such content reuse. They provide access to open and structured datasets that share information from multiple domains and can be leveraged in educational settings. Nevertheless, teachers are generally not familiar with these semantic technologies, making access and author to this data a significant challenge. This work presents a modular architecture designed to promote the reuse of learning resources that both teachers and students can use in experiences related to ubiquitous learning. The architecture integrates five managers that allow the combination of teacher-authored resources with resources from external repositories, while also providing user management and student monitoring functionalities. Two applications based on this architecture have been implemented to support learning experiences in two different educational domains.

39. KnowledgeMaps AI. Turning course materials into interactive hierarchical concept maps — with node-linked content and progress-aware learning analytics

Sergiy Tytenko, Serhii Artymovych, Vladyslav Saniuk, Ivan Feofanov, José Abreu Salas and Miquel Canal Esteve

Abstract: KnowledgeMaps AI is a practical LLM-based system that turns course materials – PDFs, slides, lecture notes – into interactive, multi-level concept maps with node-linked explanatory articles and progress-aware learning analytics. Maps give a high-level overview and support drill-down through nested child maps, letting students move from overview to focused study. LLM generation is confined to instructor-facing authoring on a serverless AWS backend, while learner sessions retrieve persisted content. We report early deployment across four courses in three institutions and three countries, with usage analytics from 178 students and perceived-usefulness surveys from three cohorts. Results indicate technical feasibility, user acceptance, and practical relevance; controlled evidence of learning gains remains future work.

40. Conversational AI Meets Mindfulness: Exploring LLMs as a Socio‑Emotional Layer in a Math Intelligent Tutoring System

Vera Rief, Mirella Hladký, Minju Yoo, Stephanie Heel, Shintaro Sato and Tomohiro Nagashima

Abstract: Intelligent Tutoring Systems (ITSs) traditionally focus their adaptive support on cognitive aspects of learning. Although effective, little is known about how such systems can be enhanced by addressing students’ emotional states. In particular, the role of mindful interventions for supporting student learning and experiences in adaptive math learning remains underexplored. We developed “”Math with Matt””, an ITS that leverages Large Language Models (LLMs) to provide both cognitive and emotional support in algebra learning. The system offers 1) an LLM-based mindful chat that delivers context-sensitive emotional support through a pedagogical agent Matt, and 2) mindful feedback and hint messages (not just evaluative) to enhance learning experiences and reduce math anxiety. We conducted a classroom study with 7th graders, comparing a Mindful version against a version with cognitive support only. Overall, the ITS reduced executive state-math anxiety and improved students’ math learning, though no significant differences emerged between the conditions. However, students with the mindfulness interventions showed higher learning efficiency and well-balanced problem-solving behavior, since they achieve a similar level of math learning with less learning time and fewer requested hints compared to the Cognitive version. Additionally, they reported that the pedagogical agent felt more supportive and caring than students in the cognitive condition. Our study demonstrates the feasibility and scalability of integrating mindfulness into ITSs through LLM-based interactions and positions LLMs as an adaptive, socio-emotional layer within cognitive math tutoring.

41. Warning About AI Fallibility Increases Help-Seeking in an Intelligent Tutoring System

Tomohiro Nagashima, Mirella Hladký and Vera Rief

Abstract: Recent work in technology-enhanced learning and Human-Computer Interaction highlights the importance of transparency and trust calibration in AI-supported learning environments as they pose a risk of hallucinations. In this study, we investigate whether a simple transparency intervention that warns students that a pedagogical agent may make mistakes affects learner behavior in a math intelligent tutoring system. We conducted a classroom experiment with 252 school students using two system versions: one including a warning message about potential system errors, and one that does not mention potential errors. Using log data, we analyzed students’ problem-solving performance data, including help-seeking behavior, error rate, and time-on-task. Results show that students who were warned about potential AI errors requested significantly more hints than those in the other condition, even though the actual system behavior was exactly the same. This finding suggests that lightweight transparency interventions can influence learners’ interaction strategies without necessarily improving or impairing immediate performance.

42. Integrating Computational Thinking through Technology‑Enhanced Learning in Special Education: An Analysis of Self-Perceived Teaching Competence and Contextual Barriers

Natalia Galbán Ojer, María Zapata Cáceres, Estefanía Martín Barroso and Marcos Román González

43. Idiographic, Aligned, and AI-Supported Learning Analytics Dashboard to Aid Student Self-Regulated Learning: Preliminary Design and Evaluation

Hesham Ahmed, Sonsoles López-Pernas, Markku Tukiainen and Mohammed Saqr

Abstract: There is a shortage of systems that are designed from the ground up to meet the ideographic principles while addressing the pedagogical and ethical misalignment. This work aims to bridge this gap by introducing a design of an aligned, idiographic, multi-source data dashboard to support students’ self-regulated learning. The app synthesizes multi-source data, namely: Learning management system (LMS) logs, sleep, screen time, and self-regulated learning (SRL) data to provide instantaneous and personalized recommendations to the students. The dashboard is supported by an AI-based companion that acts as a teacher to facilitate the interpretation of the student’s data. We also introduce a sophisticated mechanism for the context management to keep chats with the user relevant, ethical, and personal.

44. AI-Generated Feedback as a Cultural Tool in Learning and Teaching

Kseniia Makhortova

Abstract: This doctoral study examines AI-generated feedback as a mediating cultural tool in students’ writing processes and teachers’ pedagogical practices in Norwegian upper secondary schools. As part of the AI-EXCAfL project, the study compares two feedback conditions: an AI-based feedback system (FAS) and peer feedback in English and Natural Sciences classrooms. The research is grounded in cultural-historical theory, integrating Vygotsky’s concept of mediation, Galperin’s theory of phased formation of mental actions, and Ilyenkov’s dialectical concept of contradiction. Rather than measuring learning outcomes, the study traces the learning process itself, using contradiction as the central focus and unit of analysis. Data collection for the English case study is complete, comprising 1620 hours of video recordings, student draft texts across three writing cycles, and pre/post questionnaires. Preliminary interaction analysis indicates that AI feedback, based on teachers’ assessment criteria, generates categorically aligned and content-focused contradictions compared to peer feedback, where students tend to avoid critical positioning and suggest surface-level corrections. Statistical analysis is currently underway to quantify these patterns across groups and draft cycles. The study contributes an empirically grounded, theoretically enriched account of how AI feedback becomes or fails to become developmentally meaningful, with practical implications for the design of AI feedback tools and pedagogical strategies in technology-enhanced learning environments.

45. The design of pedagogical agents: Investigating learner perceptions and preferences

Julia Pöschko

Abstract: Pedagogical agents are virtual characters in digital learning environments that support student learning. Recent advances in artificial intelligence (AI) have expanded the capabilities of pedagogical agents, e.g., allowing them to flexibly and adaptively support learners as learning companions. Pedagogical agents can be used in various designs, with design decisions affecting the learning process. However, little is still known about learners’ perceptions and preferences regarding different agent designs, which play an important role for a positive learning experience. The present study investigates higher education students’ evaluations of and preferences for different designs of pedagogical agents, presented as peer-like learning companions for computational thinking. In a within-subjects vignette experiment, participants watch a series of short video vignettes in which different pedagogical agents introduce themselves as learning companions for computational thinking. The agents are varied systematically along four design factors: gender (female vs male), interaction modality (written chat vs spoken voice), level of anthropomorphism (realistic vs comic vs abstract avatar), and visual embodiment (avatar vs no avatar). After each vignette, participants rate each agent variation on learning-relevant perceptions, including trustworthiness, social presence, expected learning success, and willingness to use the agent. Additionally, learner characteristics such as gender, study discipline, computational thinking skills, self-concept, learning strategies, stereotype beliefs, and technological affinity are assessed as potential moderators. Data collection is currently ongoing. Planned analyses will examine how different agent designs shape learners’ perceptions of pedagogical agents and how individual learner characteristics moderate these relationships. The findings will help improve our understanding of how agent design, learner perceptions, and learner characteristics relate to each other, and provide empirically based recommendations for learner-centered design of pedagogical agents.

46. Beyond One-Size-Fits-All: ProFeed — A Learner-Centered Approach to Adaptive Feedback in Programming Education

Mona Münstermann

Abstract: Programming courses in higher education are characterized by high failure and dropout rates. This is partly due to the cognitive demands of learning programming and the limited ability of instructors to provide personalized feedback on a large scale. Intelligent Programming Tutors offer a potential solution, but most existing systems fail to consider individual learner characteristics when providing feedback. This paper presents ProFeed, an adaptive feedback framework for programming education based on contingent tutoring and the assistance dilemma. ProFeed aims to incorporate learners’ characteristics, such as prior knowledge and perceived self-competence, into its feedback rules.
Therefore, I conducted a Systematic Literature Review following the PRISMA guidelines to investigate how learner characteristics are currently considered in the design of adaptive feedback. The results revealed that only one of the included papers considered learner characteristics and modulated the feedback intensity accordingly. This finding highlights the disconnect between the theoretical understanding of effective feedback and its implementation in existing systems. In the next steps, these findings will inform the development of ProFeed’s initial feedback rules.

47. Designing Agentic AI System for Lecturer Decision-Making: A Hybrid Intelligence Perspective

Melike Nur Köroğlu

Abstract: Artificial Intelligence (AI) is increasingly being used to support educational decision-making in higher education. While existing AI systems provide analytics, predictions, and recommendations, limited attention has been given to how human expertise and AI capabilities should be combined within educational decision-making processes. Recent developments in Human–AI Collaboration, Hybrid Intelligence, and Agentic AI highlight the need for a deeper understanding of how complementary forms of intelligence can be designed and coordinated in educational contexts. This doctoral research investigates how Agentic AI systems can be designed to support lecturer decision-making in higher education. Adopting an Educational Design Research approach, the study combines a systematic literature review, conceptual work, socio-technical analysis, and vignette-based studies to identify and conceptualize the dimensions of Hybrid Intelligence in educational decision-making. The resulting Hybrid Intelligence framework will inform the design and development of an Agentic AI decision-support system, which will subsequently be evaluated in authentic higher education settings. The research aims to contribute both theoretical and practical insights into the design of future Human–AI collaboration in education.

48. How Students Engage with LLM Feedback Matters: Feedback Processing Time and Learning in Introductory Programming

Maciej Pankiewicz, Conrad Borchers and Ryan S. Baker

Abstract: Large Language Models (LLMs) are increasingly used to provide automated feedback in programming education. However, prior research suggests that such feedback can sometimes be too direct or too comprehensive, potentially reducing learning by over-scaffolding students. A related but understudied question is whether the time students spend processing feedback can indicate whether it is being used productively or counterproductively. This issue may be especially important for LLM-generated feedback, yet comparing it with standard compiler feedback is challenging because the two forms of feedback differ in length and processing demands, making causal interpretation difficult.
To address this issue, we analyzed data from a randomized controlled experiment in an introductory programming course that compared standard compiler error messages with compiler errors supplemented by GPT-4o-generated feedback. The LLM feedback was presented immediately after incorrect submissions as a pop-up within the learning platform. We examined effects on a final test composed of program-tracing tasks, in which students had to determine the output of pre-written code without IDE support. Using behavioral and temporal log data, we trained and validated a ridge regression model to estimate feedback engagement time and then applied this model across both experimental conditions. Predicted engagement time was a significant positive predictor of final test performance after controlling for prior knowledge, and this relationship did not differ significantly between conditions. Exploratory analyses further suggested diminishing returns at higher levels of engagement.
These findings suggest that established instructional principles also apply to LLM-based feedback: the quality of students’ engagement with assistance matters, and very rapid use may reflect less productive learning. Rather than supporting the idea that LLM feedback introduces a uniquely harmful form of “laziness,” the results point to the importance of designing systems that adaptively withhold, fade, or guide assistance in ways that encourage deeper processing.

49. Integrating Cultural Aspects into Generative AI for Higher Education: A Socio-Technical Framework for Ethical, Equitable and Contextually Responsive Practices

Patricia Khonje

Abstract: Generative Artificial Intelligence (GenAI) systems are transforming higher education but remain culturally narrow, drawing on a limited range of cultural perspectives while overlooking diverse educational traditions worldwide. In this thesis, cultural aspects refer to patterns of knowing, relating, communicating, valuing, and temporality that shape teaching and learning across communities. This limitation is particularly consequential in higher education, where learning practices, disciplinary norms, and educator–student relationships are culturally situated. As GenAI becomes embedded in educational processes, there is a risk that dominant epistemological assumptions may be reinforced at the expense of educational diversity and inclusion. This PhD research addresses this gap by developing a socio-technical framework that translates cultural dimensions of learning into actionable AI design parameters. The research explores two complementary interventions: Learning Agreements that support negotiation of expectations regarding culturally situated learning practices and AI use, and a proposed Mediator System that may enact these negotiated expectations during AI-mediated interactions. Recognizing that no single framework can fully represent cultural plurality, the study uses Hofstede’s dimensions reflexively as one heuristic while incorporating alternative perspectives through participatory stakeholder engagement. Using Design-Based Research (DBR) across three higher education contexts, flipped classrooms, project-based sustainability learning, and student-facing AI integration the research investigates how cultural dimensions can be operationalized into design parameters, develops and refines interventions, and evaluates cultural appropriateness, pedagogical alignment, ethical robustness, and educational equity through mixed methods. Expected contributions include a socio-technical framework, design principles, technical guidelines, an evaluation toolkit, and prototype interventions for culturally responsive human–AI collaboration in higher education.

50. Grounding Parent Support Design in Home Digital Learning: Observed Parent Scaffolding Behaviors in a Learner-System-Parent Framework

Yi Shang, Jingyun Wang, Xiaofei Qi and Daner Sun

Abstract: Home digital learning places substantial self-regulation demands on young children, while existing educational technologies rarely support parents as active scaffolding partners during learning. This study investigates learner-system-parent interaction in Minimarket, a self-developed early fraction learning game with an optional parent support function. The function guides parents through a three-layered workflow for identifying challenges, diagnosing specific difficulties, and selecting immediate sup-port strategies. Twenty parent-child dyads in China participated in a pilot experiment involving game-based math learning. User data, includes system logs, observation notes, parent questionnaires, and post-session interviews were collected. The study addressed three problems: uptake of the parent support function, parent scaffolding behaviors during interaction, and parents’ perceptions of the function. Findings indicate that parent-inclusive support is feasible but selectively used. Observations showed that classic scaffolding categories captured some of the parent behaviors, but additional and newly emerging behaviors were also identified in current digital home learning. Parents valued structured support but requested more concrete, task-specific guidance. The findings provide empirical grounding for future learner-system-parent support design in home digital learning.

51. Exploring Indonesian Pre-service EFL Teachers’ Cognition and Practices in GenAI-Assisted Teaching

Hasan Zainnuri

Abstract: Generative Artificial Intelligence is reshaping English as a Foreign Language teaching faster than teacher education programmes have been able to respond. In Indonesia, pre-service teachers are often left to self-directed exploration when using GenAI during micro-teaching and school practicum, the very experiences most likely to shape their long-term beliefs and practices. This PhD project investigates how pre-service EFL teachers’ cognition of GenAI-assisted teaching develops longitudinally across these two phases, and how that cognition relates to what teachers actually do in the classroom. A mixed methods longitudinal multiple case study design was used, combining a survey of 94 participants with in-depth case analysis of six, drawing on AI-TPACK and TAM questionnaires, semi-structured interviews, and Stimulated Recall Interviews. Findings from the first completed case show that cognition trajectories are uneven rather than linear, shaped by the interaction of prior GenAI literacy, actual classroom usage, and student responses, mediated through what this study identifies as the supervision triad of teacher, mentor, and university supervisor. These findings extend the language teacher cognition framework to incorporate GenAI usage practices and the contextual factors that support or hinder cognitive change, pointing toward interventions that target the whole supervision triad rather than the individual teacher alone.

52. Mapping Teacher Agentic Properties to Orchestration Activities in AI-Augmented Classrooms

Víctor Alonso-Prieto, Vincent Aleven, Yannis Dimitriadis, Qianru Lyu, Juan Ignacio Asensio-Pérez, Ken Holstein and Qiao Jin

Abstract: Preserving and fostering teacher agency is viewed as a foundation for effective and trustworthy AI deployment in educational settings. However, agency is a difficult concept to operationalize, especially when AI tools affect teachers’ cognitive processes, hindering the identification and interpretation of agency in instructional decision-making. To tackle this issue, we propose a new conceptualization of teacher agency that considers the critical dimension of human-AI interaction. Our new framework builds upon existing frameworks to combine different theorizations of agency with orchestration activities. This paper reports the results from a case study in which this novel conceptualization informed the research design, crafted the interview script, and provided a high-level structure for interpreting the findings. The study involved a pilot implementation of an AI-supported Mixed Reality (MR) tool to augment teachers’ perceptions in settings where middle-school students work with an Intelligent Tutoring System. Three teachers were studied through classroom observations and interviews. Our findings illustrate how the framework supports the interpretation of teacher agency unfolding in dynamic classroom scenarios and suggest that teachers perceive the MR tool as a valuable companion for classroom regulation, supporting their sense of productivity. Articulating the study of agency through orchestration activities explicitly augmented by AI has the potential to account for how agency manifests in AI classrooms. Consequently, new research directions may emerge, positioning agency as a core design principle in TEL, aiming to facilitate mechanisms that empower teachers to capitalize on augmented perception by exercising autonomous judgments leading to more intentional actions and enhanced self-efficacy.

Cookie	Duration	Description
cookielawinfo-checkbox-analytics	11 months	This cookie is set by GDPR Cookie Consent plugin. The cookie is used to store the user consent for the cookies in the category "Analytics".
cookielawinfo-checkbox-functional	11 months	The cookie is set by GDPR cookie consent to record the user consent for the cookies in the category "Functional".
cookielawinfo-checkbox-necessary	11 months	This cookie is set by GDPR Cookie Consent plugin. The cookies is used to store the user consent for the cookies in the category "Necessary".
cookielawinfo-checkbox-others	11 months	This cookie is set by GDPR Cookie Consent plugin. The cookie is used to store the user consent for the cookies in the category "Other.
cookielawinfo-checkbox-performance	11 months	This cookie is set by GDPR Cookie Consent plugin. The cookie is used to store the user consent for the cookies in the category "Performance".
viewed_cookie_policy	11 months	The cookie is set by the GDPR Cookie Consent plugin and is used to store whether or not user has consented to the use of cookies. It does not store any personal data.