Providing individualized support in introductory programming is challenging for teachers in secondary classrooms. Although LLMs can effectively deliver automated feedback in this context, they may provide excessive assistance that hinders learning and lack access to the full context of exploratory environments such as serious games. This study investigates the impact of integrating a pedagogically driven LLM-based digital assistant, which integrates distinct constructivist constraints, into Pyrates, a Python programming game for high schools. In a field user study (N = 241), we found that the assistant was positively perceived by students, supported in-game progression, and significantly reduced teacher interventions without negatively affecting learning. Interaction trace analyses revealed distinct post-feedback strategies, advocating for a tight control of LLM content generation. These findings highlight the feasibility and value of embedding pedagogical principles into LLMs to reduce teacher workload while influencing student strategies and maintaining their learning outcomes in authentic K-12 programming courses.

Educational videos are a cornerstone of remote and blended learning. However, learners' fluctuating attention remains a significant barrier to effective information retention. Prior research has attempted to mitigate this by detecting and reacting to attention loss at runtime using eye tracking. Such detection has been based so far on classical machine learning classifiers trained on engineered features, such as summary statistics over learners' fixations and saccades. These methods have struggled to capture the complex, temporal nature of learner engagement, thus exhibiting moderate prediction performance. In this study, we aim to advance the detection of attention by shifting from standard engineered features to a multimodal foundation models. Using an educational eye-tracking dataset (N = 70), we investigate a novel methodology that utilizes a Vision-Language Model (VLM) to analyze video content directly with superimposed gaze data. This approach aims to leverage the semantic reasoning capabilities of foundation models to contextualize learner focus within the video stream. We evaluate the performance of this VLM-based approach using several prompting strategies with Gemini 3, but ultimately found that none of them could outperform statistical baselines. Our results provide new insights into the limitations of using VLMs for real-time educational diagnostics.

Predictive student models are increasingly used in learning environments. However, due to the rising social impact of their usage, it is now all the more important for these models to be both sufficiently accurate and fair in their predictions. To evaluate algorithmic fairness, a new metric has been developed in education, namely the Model Absolute Density Distance (MADD). This metric enables us to measure how different a predictive model behaves regarding two groups of students, in order to quantify its algorithmic unfairness. In this paper, we thus develop a post-processing method based on this metric, that aims at improving the fairness while preserving the accuracy of relevant predictive models' results. We experiment with our approach on the task of predicting student success in an online course, using both simulated and real-world educational data, and obtain successful results. Our source code and data are in open access at https://github.com/melinaverger/MADD .

Research has shown that serious games can effectively support programming learning in K-12 education. However, some students may have difficulties learning with such games and, in those cases, can benefit from dedicated support. Unfortunately, K-12 teachers cannot always provide this support in a timely manner, especially due to a lack of time in crowded high school classes. Thus, to facilitate the large-scale use of programming serious games in high school courses, we explore the feasibility of augmenting such games with AI-driven adaptive feedback designed to address some of the learners' difficulties while reducing teachers' burden. Specifically, we investigate the design of an AI-driven feedback policy that can predict, during interaction with a serious game, what type of formative feedback is suitable to address students' difficulties. To build this policy, we conducted a user study in regular 10th grade Python classes to collect labels on how experienced teachers (N = 7) support their students (N = 215). We trained machine learning classifiers to predict these labels using students' interaction behaviours and code. Our results show that teachers extensively use several feedback types and that Random Forest classifiers can significantly outperform relevant baselines in predicting them, achieving a F1 score, precision, and recall of up to 0.70. We also shed light on the behaviours that best explain these predictions, and found that features related to both the students' code and their interaction with the game elements relevant to the current game level matter. We then discuss the implications of our work for designing personalized feedback at scale in programming serious games.

The expense and logistics of organizing experiments to train and evaluate teaching policies, as well as the potential negative impacts of these policies on the very first students, are significant challenges in the field of education. In this paper, we explore the feasibility of using offline reinforcement learning (RL) to learn an adaptive feedback policy that improves student progress in Pyrates, a programming platform. Leveraging an existing dataset of teacher-provided feedback and student interactions and codes, we have developed an offline RL model capable of learning an optimal feedback policy without direct interaction with students. The trained policy is then evaluated to assess its effectiveness in maximizing student progress and its suitability for online deployment in real-world educational settings. Our evaluation yields promising results, demonstrating that offline-trained policies can significantly enhance student progress, highlighting their potential for scalable deployment in programming platforms. However, challenges remain in ensuring robustness and adaptability when transitioning to online use.

This paper presents a scoping review of literature on using generative artificial intelligence (GenAI) in computer science (CS) education at the higher education (HE) level. This review's primary research question is: "How is GenAI utilized in the teaching and learning of CS in HE?" A total of 890 articles were initially identified following a protocol based on PRISMA guidelines. After the selection screening, 38 articles were included for quantitative and qualitative analysis. Students mainly use GenAI to assist with programming tasks. However, there is a lack of research for developing effective pedagogical strategies to fully integrate GenAI into CS education. While opinions on its effectiveness in HE are mixed, students' growing enthusiasm for GenAI stresses the critical need for research on pedagogical approaches to effectively handle its integration and facilitate the current educational paradigm shift. This shift calls for proactive pedagogical strategies and research to adapt teaching methods to an AI-driven context. Strategic GenAI use and regulation will equip new CS learners to thrive in the evolving tech landscape.

This paper focuses on analyzing a 3D virtual educational escape game, which helps players solve puzzles by offering hints automatically. The study aimed to determine whether these automated hints improved player retention and progression, and if these effects are influenced by the players' gaming skills. We present our results on two instances of an escape game, which have the same game structure but different learning content, the first focusing on accessibility and inclusive education (ACCESS version) and the second on computer science (CS version). Although the structure of the two versions is identical, we found that the results differ between them. The automatic hint system was very effective in the ACCESS version, improving both retention and progression. In contrast, it is less effective in the CS version, where the complexity of the puzzles might have hindered progression.

This study evaluates the impact of an adaptive feedback system in Pyrates, a programming serious game designed to ease the transition from block-based to text-based programming in high school classes. The adaptive feedback system was implemented to support student learning and lessen teachers' intervention workload in the classroom. To assess its effectiveness, a field user study was conducted with 190 high school students across two institutions. Results show that students progressed significantly further in the game when using the adaptive feedback system, as compared to playing without feedback, although it did not affect learning gains. We discuss the implications of these results for the design of adaptive feedback in programming serious games.

Educational videos are widely used in remote and blended learning. However, learners' attention often fluctuates while watching, which can hinder their retention of key information. This, in turn, may impact their overall learning outcomes. Detecting when learners lose attention or fail to memorize key elements of a video could help address these challenges-for example, by enabling adaptive support that enhances engagement and retention, bridging the gap between passive video consumption and active learning. Such automated detection could also provide valuable insights to instructors about when attention and retention drop in their videos. In this study, we explore how to detect learners' attention and retention while they watch an educational video in a blended course on green chemistry, using eye-tracking data. To achieve this, we develop machine learning classifiers that analyze eye movements, pupil dilation, eye-screen distance, and attention to dynamically tracked areas of interest (AOIs). We investigate different strategies for fusing these types of information and find that dynamic AOIs can significantly improve ML predictions, albeit with moderate performance.

This workshop aims to provide a venue for researchers to discuss latest work on AI-based support for Exploratory Learning Environments (ELE). It builds on a series of workshops in this area and latest advancements from the organising committee but also several colleagues who have indicated an interest in this area.

While machine learning (ML) has been extensively used in Massive Open Online Courses (MOOCs) to predict whether learners are at risk of dropping-out or failing, very few work has investigated the bias or possible unfairness of the predictions generated by these models. This is however important, because MOOCs typically engage very diverse audiences worldwide, and it is unsure whether the existing ML models will generate fair predictions to all learners. In this paper, we explore the fairness of ML models meant to predict course completion in a MOOC mostly offered in Europe an Africa. To do so, we leverage and compare ABROCA and MADD, two fairness metrics that have been proposed specifically in education. Our results show that some ML models are more likely to generate unfair predictions than others. Even in the fairest models, we found biases in their predictions related to how the learners' enrolled as well as their country, gender, age and job status. These biases are particularly detrimental to African learners, which is a key finding as they are an understudied population in AI fairness analysis in education.

Algorithmic bias is a major issue in machine learning models in educational contexts. However, it has not yet been studied thoroughly in Asian learning contexts, and only limited work has considered algorithmic bias based on regional (sub-national) background. As a step towards addressing this gap, this paper examines the population of 5,986 students at a large university in the Philippines, investigating algorithmic bias based on students' regional background. The university used the Canvas learning management system (LMS) in its online courses across a broad range of domains. Over the period of three semesters, we collected 48.7 million log records of the students' activity in Canvas. We used these logs to train binary classification models that predict student grades from the LMS activity. The best-performing model reached AUC of 0.75 and weighted F1-score of 0.79. Subsequently, we examined the data for bias based on students' region. Evaluation using three metrics: AUC, weighted F1-score, and MADD showed consistent results across all demographic groups. Thus, no unfairness was observed against a particular student group in the grade predictions.

Adaptive support can help learners in Open-Ended Learning Environments (OELEs), where the free-form nature of the interaction can be confusing to students. In this paper, we design and evaluate an Intelligent Pedagogical Agent (IPA) for an OELE designed to foster Computational Thinking (CT). Specifically, we design help interventions for an in-the-wild scenario where students interact with the OELE in an unmonitored, self-directed manner. We build a student model by extracting meaningful student behaviors on real-world interaction data obtained during interaction in online classrooms and including expert insights. We show that these student models perform better than a baseline and have the potential for adaptive support in self-directed interaction with the OELE. We design an IPA with the help of teachers, leveraging the student behaviors extracted from data. Lastly, we get insights into the value of these help interventions by empirically evaluating the IPA in a formal user study.

There is extensive evidence that AI-based educational technology can effectively provide personalized support to students’ problem solving in many domains. In contrast, there has been limited work on AI-based environments that support exploratory activities, such as learning from interactive simulations or playing educational games. These activities are becoming increasingly widespread, especially in MOOCs (Massive Open On-Line Courses) and in other forms of self-directed learning, because they can foster motivation and grounded skills acquisition. However, not all learners possess the reasoning and study skills to fully benefit from these activities, calling for environments that can provide adaptive support for open-ended exploratory learning. Building such tools is challenging because it requires student models that assess behaviors for which there is no clear definition of correctness, as well as skills beyond those traditionally addressed by student models for problem solving. This chapter will review current advances of student modeling research in this direction.

The eLearning specification xAPI, which employs a shared format for receiving and transmitting data, is used to collect data about the diverse range of experiences within online learning activities, thereby enabling the exchange of knowledge between multiple systems. This paper presents Logstore xAPI, a plugin that emits Moodle events as xAPI statements, allowing a modular, interoperable, performing and secure way of logging user interactions and learning experiences, and send them to a Learning Record Store (LRS) to be stored and further analysed. We describe all phases of mapping Moodle events to statements and storing them in the LRS, along with the issues we encountered and our solutions.

Predictive student models are increasingly used in learning environments due to their ability to enhance educational outcomes and support stakeholders in making informed decisions. However, predictive models can be biased and produce unfair outcomes, leading to potential discrimination against some students and possible harmful long-term implications. This has prompted research on fairness metrics meant to capture and quantify such biases. Nonetheless, so far, existing fairness metrics used in education are predictive performance-oriented, focusing on assessing biased outcomes across groups of students, without considering the behaviors of the models nor the severity of the biases in the outcomes. Therefore, we propose a novel metric, the Model Absolute Density Distance (MADD), to analyze models' discriminatory behaviors independently from their predictive performance. We also provide a complementary visualization-based analysis to enable fine-grained human assessment of how the models discriminate between groups of students. We evaluate our approach on the common task of predicting student success in online courses, using several common predictive classification models on an open educational dataset. We also compare our metric to the only predictive performance-oriented fairness metric developed in education, ABROCA. Results on this dataset show that: (1) fair predictive performance does not guarantee fair models' behaviors and thus fair outcomes, (2) there is no direct relationship between data bias and predictive performance bias nor discriminatory behaviors bias, and (3) trained on the same data, models exhibit different discriminatory behaviors, according to different sensitive features too. We thus recommend using the MADD on models that show satisfying predictive performance, to gain a finer-grained understanding on how they behave and to refine models selection and their usage.