Collaborative assignments are a cornerstone of programming education. Effective collaboration during a programming project depends on the formation of reasonably accurate beliefs about how each partner works. Generative AI tools, now widely used by undergraduate students, have introduced a consequential and largely invisible new dimension into collaboration: each student's use of AI. When partners collaborate remotely, they interpret partners' ability and effort through their code. This raises the question of how accurately students perceive each other's AI use in collaborations, and if a misalignment in these perceptions relates to team performance. To address this question, we conducted a three-wave longitudinal study of 103 student pairs in an introductory software engineering course. We found that greater misalignment between partners' beliefs about each other's AI use early in the project was associated with lower final project scores. The effect of such misaligned perceptions is the strongest in teams with lower prior programming performance, suggesting that low performing students pay a higher cost of misaligned perceptions. The perception misalignment does not consistently decrease through face-to-face pair-programming sessions. This suggests that ways to foster transparency may be needed to support student teams in collaborative programming.

As students learn and practice new skills in university courses, their behaviour can change in response to competing demands and increasing content complexity. However, most metrics used to evaluate study behaviour focus on the number or sequence of activities rather than on the change of behaviour. To address this, we replicate and extend a complex dynamical systems approach to characterise recurrence in behavioural patterns and whether it changes.Using assessment logs from 1362 students in the first 5 weeks of a semester-long programming course, we examine whether changes in the patterns of formative assessment submissions can differentiate student sub-groups and predict their performance. We identify three student profiles of behavioural change. We find that higher entropy of recurrence in assessment submission patterns is associated with better performance, and that changes in this entropy signal upcoming changes in performance. We also show that higher entropy of recurrence is associated with greater timeliness of submissions. Finally, we evaluate the predictive value of early behavioural patterns and find that while student profiles of change do not outperform conventional predictive metrics, they offer complementary insights that can enable timely interpretations of student data and inform interventions. Overall, our findings extend the generalisability of behavioural metrics based on complex dynamical systems by demonstrating consistent patterns across courses, LMS types and data sources.

Artificial intelligence increasingly shapes higher education, yet most evaluations focus on average learning outcomes and overlook differences across student groups such as gender. This paper presents a secondary analysis of a controlled experimental study comparing a scaffolded, institutionally embedded tutoring system (IRIS), an unrestricted conversational AI tool (ChatGPT), and a no-AI condition. The original study found that AI support improved task performance but did not consistently improve learning outcomes. This analysis extends the model by testing whether effects differ by gender.An ANCOVA on posttest scores controlling for pretest revealed a small but significant Condition × Gender interaction (F(2,383) = 3.55, p = .030; f2 = .019), driven by the ChatGPT condition: female students showed higher adjusted posttest scores than the no-AI control (d ≈ +0.28), whereas no comparable effect emerged for male students (d ≈−0.19) or in the IRIS condition. The pattern held across six sensitivity analyses. No comparable interaction emerged for exercise performance, intrinsic motivation, or cognitive load.These findings suggest that AI-supported learning environments are not uniform in their effects across student groups. The two AI tools differ along several dimensions simultaneously, so the present design cannot isolate which dimension drives the moderation; this analysis adopts "how AI structures access to help" as an interpretive framing motivated by prior work on stereotype threat and ambient belonging rather than as an identified mechanism.

Learning at scale often requires domain-specific automation such as assessment and feedback. An organization locked in to a general learning platform without these specialist automations limits its pedagogical offering. An ecosystem of interoperable, platform-agnostic microservices for domain-specific automation would solve this problem. To develop an effective ecosystem, a standard interface (API) for education microservices is required. We propose an initial specification for a standard, platform-independent API for educational microservices, μEd. The API integrates functionality from existing systems in use at four institutions, which are adopting the new API. The API is initially specified for automation of feedback, assessment, and educational chatbots, with further service types planned. The API specification provided here enables the development of an ecosystem of education microservices that will facilitate automation in more domains, to more users, providing a richer learning experience in a wide range of disciplines.

University courses with hundreds of students have become common, particularly during early years of university studies. The sheer scale of these courses limits traditional instruction, shifting it towards a one-to-many mode of delivery. This shift reduces student-instructor interaction and tailored instructor feedback which are crucial for student success. Automated feedback systems allow scaling feedback, but they often reduce instructor contributions to student learning. This paper investigates how emerging technologies can support, rather than replace, instructors in tailoring their teaching and feedback to identify and correct student knowledge gaps at scale. To address this challenge, the paper introduces a novel technological solution: the Knowledge Gaps to Mastery (KG2M) approach. KG2M combines discussion forum data with course-specific content and leverages large language models (LLMs) and Retrieval-Augmented Generation (RAG) for the dual purpose of identifying prevalent class-level knowledge gaps and transforming them into targeted learning activities and formative assessments. The approach was deployed across three computer science courses with a combined enrollment of 1,355 students and evaluated through semi-structured interviews with five instructors. Results indicate that instructors found the tool intuitive and pedagogically valuable, particularly for surfacing knowledge gaps and generating actionable teaching insights. The paper reports on the tool, the evaluation, and the current limitations of the approach that emerged during instructor evaluation.

Introduction: Generative AI is reshaping programming education, yet its effects on conceptual learning, intrinsic motivation, and cognitive load remain unclear. This study tests whether assistance deepens understanding or primarily boosts task completion, and how scaffolded versus answer-giving designs matter. Objectives: This study compares performance, learning, cognitive load, frustration, and motivation across three AI support types, and examines students' perceptions. Methods: A three-arm randomized controlled trial was conducted in an introductory programming (CS1) course at TUM (N=275). Participants completed a 90-minute exercise on concurrency, implementing a parallel sum with threading in one of three conditions: (1) Iris, a scaffolded tutor providing calibrated hints while withholding full solutions; (2) ChatGPT, unrestricted assistance that can provide complete solutions; (3) no-AI control using traditional web resources. Pre-and post-knowledge tests and a code comprehension task measured learning, while auto-graded test coverage measured performance. Validated scales captured intrinsic, germane, and extraneous cognitive load, frustration, and intrinsic motivation. Results: Both AI groups achieved substantially higher exercise scores than the control group, with distinct distri butions: ChatGPT users clustered at high scores, control participants at low scores, and Iris users spread across the full range. Despite these performance gains, neither AI condition produced greater pre-post knowledge gains or code-comprehension advantages. Both AI groups reported lower frustration and reduced extraneous and germane load than the control group, while intrinsic load did not differ. Only Iris increased intrinsic motivation. Students rated ChatGPT as easier to use and more helpful. Conclusion: In this setting, generative AI acted primarily as a performance aid rather than a learning enhancer. Scaffolded, hint-first design preserved motivational benefits, whereas AI providing unrestricted solutions en couraged a "comfort trap" where students' preferences misaligned with pedagogical effectiveness. These findings motivate scaffolded AI integration and assessment designs resilient to environments where performance no longer reliably tracks understanding.

Learning Analytics Dashboards (LADs) have shown limited impact on student outcomes, often functioning as static visualizations. In this paper, we present a case study reimagining LADs as interactive tools that promote student engagement. Namely, we extend a conventional LAD with a Large Language Model (LLM)-powered pedagogical agent and an interactive Judgement of Learning (JoL) feature to support students’ awareness of their study progress. The pedagogical agent engaged students in conversations about their learning data, while the JoL feature required self-assessment before viewing system metrics, encouraging learners’ calibration of their judgement. This interactive LAD (ILAD) was implemented in a university programming course with three randomly assigned conditions: no agent, a “telling” agent providing information about learner data, and an “eliciting” agent asking questions about learner data. The case study reports on the data collected over five weeks from thirty students paid to regularly use the ILAD within their computer science course. Data analysis from this small sample showed that the students in the “elicit” condition engaged in more reflection and more accurately judged their own mastery. This study highlights ILADs’ potential to foster student engagement with learning data and improve metacognitive outcomes, offering new directions for learning dashboard design.

The increasing demand for programmers has led to a surge in participants in programming courses, making it increasingly challenging for instructors to assess student code manually. As a result, automated programming assessment systems (APASs) have been developed to streamline this process. These APASs support lecturers by managing and evaluating student programming exercises at scale. However, these tools often do not provide feature-rich online editors compared to their traditional integrated development environments (IDEs) counterparts. This absence of key features, such as syntax highlighting and autocompletion, can negatively impact the learning experience, as these tools are crucial for effective coding practice. To address this gap, this research contributes to the field of programming education by extracting and defining requirements for an online IDE in an educational context and presenting a prototypical implementation of an open-source solution for a scalable and secure online IDE. The usability of the new online IDE was assessed using the Technology Acceptance Model (TAM), gathering feedback from 27 first-year students through a structured survey. In addition to these qualitative insights, quantitative measures such as memory (RAM) usage were evaluated to determine the efficiency and scalability of the tool under varying usage conditions.

In the contemporary educational landscape, particularly in large classroom settings, discussion forums have become a crucial tool for promoting interaction and addressing student queries. These forums foster a collaborative learning environment where students engage with both the teaching team and their peers. However, the sheer volume of content generated in these forums poses two significant interconnected challenges: How can we effectively identify common misunderstandings that arise in student discussions? And once identified, how can instructors use these insights to address them effectively? This paper explores the approach to integrating large language models (LLMs) and Retrieval-Augmented Generation (RAG) to tackle these challenges. We then demonstrate the approach Misunderstanding to Mastery (M2M) with authentic data from three computer science courses, involving 1355 students with 2878 unique posts, followed by an evaluation with five instructors teaching these courses. Results show that instructors found the approach promising and valuable for teaching, effectively identifying misunderstandings and generating actionable insights. Instructors highlighted the need for more fine-grained groupings, clearer metrics, validation of the created resources, and ethical considerations around data anonymity.

In large university courses, students face challenges in mastering programming skills, which underlines the need for effective pedagogical support. The Artemis learning platform has been providing interactive learning through automated, individual feedback since 2016. It also offers fully personalized support with detailed explanations based on generative AI. The general availability of large language models enabled the creation of IRIS, a virtual assistant designed to offer context-sensitive support. Deployed as part of Artemis, IRIS is accessible to thousands of students. This presentation highlights the challenges of developing AI assistance in education and motivates ongoing research focused on analyzing student behavior and evaluating the impact on their learning experience.

Auto-graded programming exercises involve multiple interconnected artifacts including problem statements, code templates, reference solutions, and test suites. Inconsistencies between these artifacts create extraneous cognitive load, forcing students to reconcile contradictory information rather than focus on learning objectives. Educators must manually detect such inconsistencies, which is time-consuming and error-prone. We present a theoretically grounded approach to automated consistency verification for multi-artifact programming exercises, building on Mayer's Coherence Principle and Biggs' Constructive Alignment. The methodology combines a novel educational artifact ontology with a specialized Large Language Model (LLM) pipeline detecting Structural and Semantic inconsistencies across heterogeneous exercise components. The ontology defines five consistency categories (Structural, Semantic, Assessment, Temporal, Scope), and this release operationalizes the first two while reserving the remaining three for future instantiation. We evaluate on 91 perturbed variants from three Java exercises that contain 93 annotated issues across six sub-categories. The reference o4-mini configuration yields 63% precision, 91% recall, F1 0.75, and span F1 0.68, recovering nine in ten inconsistencies with accurate spans; Structural mismatches peak at F1 0.87, whereas Semantic naming remains the main source of noise at F1 0.72. Grok 3 Mini halves latency to 14.3 s and cost to $0.006 per run while retaining F1 0.63. We release the PECV-bench replication package to support reproducible and extensible baselines.

Timely and individualized feedback is essential for students' learning progress and motivation, yet providing such feedback has become increasingly challenging due to growing student numbers. This has resulted in a time-consuming, repetitive, and often manual task for educators, contributing to a high workload. This paper presents DAFeeD, an LLM-based approach for automated feedback on student submissions across various exercise domains. The defined feedback process enables interactive learning by allowing students to submit solutions multiple times and automatically receive iterative LLM feedback on their submission attempts before deadlines. By incorporating task details, grading criteria, student solutions, and custom instructions into the prompt, DAFeeD provides clear, personalized, and pedagogically meaningful feedback to support continuous improvement. To evaluate the feedback process, we implemented DAFeeD in an open-source reference implementation integrated into the learning platform Artemis. A controlled study with students working on a programming task in a supervised environment showed that students found the feedback relevant and beneficial. They reported feeling more comfortable and willing to request automated feedback due to its convenience and immediacy. Additionally, deploying DAFeeD in a software engineering course with 450 students demonstrated improvements in student performance and encouraged iterative refinement through multiple submissions. These findings highlight DAFeeD's potential to enhance feedback processes in computing education, improving both learning efficiency and student outcomes.

The ability of large language models (LLMs) to process, interpret, and comprehend vast amounts of heterogeneous data presents a significant opportunity to enhance data-driven care delivery. However, the sensitive nature of protected health information (PHI) raises concerns about data privacy and trust in remote LLM platforms. Additionally, the cost of cloud-based artificial intelligence (AI) services remains a barrier to widespread adoption. To address these challenges, we propose shifting the LLM execution environment from centralized, opaque cloud providers to a decentralized and dynamic fog computing architecture. By running open-weight LLMs in more trusted environments, such as a user's edge device or a fog layer within a local network, we aim to mitigate the privacy, trust, and financial concerns associated with cloud-based LLMs. We introduce SpeziLLM, an open-source framework designed to streamline LLM execution across multiple layers, facilitating seamless integration into digital health applications. To demonstrate its versatility, we showcase SpeziLLM across six digital health applications, highlighting its broad applicability in various healthcare settings.

Soft skills are critical for academic and professional success, but are often neglected in early-stage technical curricula. This paper presents a semi-isolated teaching intervention aimed at fostering study ability and key soft skills-communication, collaboration, and project management-among first-year computer science students. The elective seminar Soft Skills and Tools for Studies and Career in IT was taught alongside a mandatory team-based programming course. We analyze project outcomes and student experiences across three cohorts across three groups: students who attended the seminar, students who teamed up with a seminar attendee, and students with no exposure to the seminar. Results show that seminar participants performed significantly better in individual presentations and team projects. Qualitative feedback further indicates improved team dynamics and study preparedness. Although self-assessed collaboration and communication did not reach statistical significance, consistent trends suggest that early soft skills training enhances academic integration. Based on these results, we recommend embedding such interventions early in technical study programs to support the transition into university life.

Large programming courses face critical challenges in providing personalized support at scale. Educators have responded by developing specialized AI tutors, yet empirical research comparing their effectiveness to general-purpose tools and traditional support remains scarce. We conducted an exploratory randomized, between-subjects mixed-methods study (N=33) in which students implemented the Burrows-Wheeler Transform under one of three conditions: a context-aware AI tutor (Iris), a general-purpose chatbot (ChatGPT), or no AI support. Quantitative analyses detected no significant differences in learning gains, completion time, or code accuracy, while qualitative interviews unveiled key insights: time pressure dominated tool selection, with students prioritizing efficiency over learning under stress; context-aware guidance was universally appreciated; students showed polarized scaffolding preferences; and ChatGPT users sought external verification more than Iris users. Over-reliance concerns were prevalent, with ChatGPT users expressing stronger concerns about this issue post-intervention. These findings indicate that tool design and situational context strongly shape AI adoption even when performance metrics converge, underscoring the need for larger comparative trials to determine which learners benefit from different AI supports and how to integrate them responsibly.

This study investigates teamwork dynamics in student software development projects through a mixed-method approach combining quantitative analysis of GitLab commit logs and qualitative survey data. We analyzed individual contributions across six project phases, comparing self-reported and actual contributions to measure discrepancies. Additionally, a survey captured insights on team leadership, conflict resolution, communication practices, and workload perceptions. Findings reveal that teams with minimal contribution discrepancies achieved higher project grades and exam pass rates. In contrast, teams with more significant discrepancies experienced lower performance, potentially due to role clarity and communication issues. These results underscore the value of shared leadership, structured conflict resolution, and regular feedback in fostering effective teamwork, offering educators strategies to enhance collaboration in software engineering education through self-reflection and balanced workload allocation.

The rapid adoption of AI powered coding assistants like ChatGPT and other coding copilots is transforming programming education, raising questions about assessment practices, academic integrity, and skill development. As educators seek alternatives to traditional grading methods susceptible to AI enabled plagiarism, structured peer assessment could be a promising strategy. This paper presents an empirical study of a rubric based, anonymized peer review process implemented in a large introductory programming course. Students evaluated each other's final projects (2D game), and their assessments were compared to instructor grades using correlation, mean absolute error, and root mean square error (RMSE). Additionally, reflective surveys from 47 teams captured student perceptions of fairness, grading behavior, and preferences regarding grade aggregation. Results show that peer review can approximate instructor evaluation with moderate accuracy and foster student engagement, evaluative thinking, and interest in providing good feedback to their peers. We discuss these findings for designing scalable, trustworthy peer assessment systems to face the age of AI assisted coding.