AI4HE 2026 Abstracts

Area 1 - Artificial Intelligence for Higher Education
Full Papers Short Papers

Area 1 - Artificial Intelligence for Higher Education

Full Papers
Paper Nr: 7
Title:

Generative AI in Computer Science Education: Insights from an Exploratory Study of Students, Educators, and Industry Professionals
Authors:

Laura Diana Cernău, Laura Dioşan and Camelia Şerban
Abstract: This study offers preliminary insights into the adoption of Generative Artificial Intelligence (GenAI) tools in computer science (CS) education, drawing on responses from 202 university-level computer science students, 16 educators, and 6 industry professionals. Employing a three-dimensional approach, the study delivers a nuanced, cross-stakeholder perspective on current GenAI utilisation and perceptions within CS educational contexts. The findings indicate a high prevalence of GenAI tools usage among students, who primarily view these as productivity enhancers, while also acknowledging their limitations and ethical challenges. Both students and educators emphasise the necessity for structured guidance on responsible tool usage. Educators further assert that generative AI should serve as a support mechanism rather than replace foundational knowledge, and they advocate establishing formal usage guidelines. Industry professionals support this view, recognising the critical role of GenAI in modern software development and stressing the importance of students acquiring core computer science competencies. The study concludes that, although GenAI presents considerable opportunities, there is a significant risk of overreliance that could hinder critical thinking. Consequently, explicit and intentional institutional guidance is required to ensure the effective and ethical integration of GenAI tools into CS education. Although this exploratory, context-bound study does not seek to generalise its conclusions, it underscores the urgency for institutions to develop and implement robust strategies supporting responsible GenAI adoption. Further, larger-scale research is needed to determine how educational practices can optimise the benefits of GenAI and mitigate associated risks in a rapidly evolving technological environment.
Download
Paper Nr: 8
Title:

Design and Evaluation of a Historical Conversational Agent Embodying Carlo Pisacane
Authors:

Fabio Clarizia, Pasquale Esposito, Annarita Giunto and Rocco Loffredo
Abstract: This paper presents a historical conversational agent embodying Carlo Pisacane. Conversational agents based on Large Language Models offer new opportunities for education and cultural heritage. However, they suffer from inaccuracies and temporal inconsistencies. To address this, we adopt a Retrieval-Augmented Generation (RAG) approach that grounds responses in a historical documentary corpus. We propose a framework combining document grounding, temporal constraints, and persona modeling. Multiple language models were tested to evaluate narrative coherence, historical accuracy, and mitigation of hallucinations. Results show that medium-sized models provide the best trade-off between generative quality and factual fidelity. A user study (n=52) reports high usability, engagement, and educational value. These findings demonstrate the potential of RAG-based historical agents to enhance learning and cultural engagement.
Download
Paper Nr: 13
Title:

PathMod: A Multi-Agent Pedagogical Tool for Learning Pathway Generation and Knowledge Dependency Visualization
Authors:

Nipun Dhokne and Suman Kundu
Abstract: Traditional learning management systems present educational content in fixed, linear sequences that ignore inter-dependencies between learning concepts within the content. Students often encounter advanced topics before mastering foundational concepts, as educators lack automated tools to analyze and restructure learning materials based on knowledge dependencies. Manual curriculum organization is time-consuming, and prerequisite violations are common. Current systems cannot adapt to diverse educational content or provide dependency-aware learning pathways at scale. PathMod addresses these challenges through a seven-agent system that automatically transforms educational PDF slides into prerequisite-aware learning modules. The system employs a three-stage pipeline: first, the Content Extractor and Concept Extractor identify key concepts and definitions from slide content; second, the Dependency Analyzer analyzes prerequisite relationships to construct knowledge dependency graphs (DAGs); and third, the Module Organizer implements a two-phase approach of topological level-based partitioning followed by smart balancing with concept movability classification to achieve balanced distribution without breaking dependencies. The Learning Path Generator creates meaningful learning sequences within each module, Module Summarizer produces comprehensive educational summaries, and Evaluator Agent performs independent cross-model quality assessment of both concept extraction and dependency graph construction. PathMod enables educators to rapidly create structured, adaptive learning experiences from existing materials while ensuring students encounter concepts in an optimal, rationale enriched learning sequence that respects prerequisite ordering and maintains semantic coherence.
Download
Paper Nr: 14
Title:

PrereqGen: An Agentic Multi-Stage Framework for Pedagogically Grounded Prerequisite Question Generation
Authors:

Rohit Kumar Goyal and Suman Kundu
Abstract: Assessing student prerequisite knowledge before introducing advanced topics is fundamental to effective pedagogy, yet manually crafting such assessments remains time-intensive and requires substantial curricular expertise. Existing automated question generation (AQG) systems typically generate questions from instructional content. They rarely model prerequisite relationships between concepts and often lack grounding in external knowledge sources. Consequently, current AQG approaches fail to support prerequisite assessment generation, leaving an important gap between automated question generation research and instructional requirements. We present PrereqGen, a novel end-to-end agentic pipeline for generating prerequisite assessments from educational documents. PrereqGen utilizes a multi-agent architecture in which each agent operates as a domain-specialized expert: a Concept Extractor Agent performs semantic analysis of educational PDFs to identify the primary topic and key concepts, a Prerequisite Finder Agent identifies foundational concept dependencies, and a Question Generator Agent produces pedagogically grounded assessment items enriched with Wikipedia context through retrieval-augmented generation. The generated assessments are automatically published to Google Classroom, bridging the gap between automated question generation research and practical classroom deployment. Evaluation on 75 computer science course materials and 1,500 generated questions demonstrates that PrereqGen achieves an average overall quality score of 2.64/3.0 (3-point Likert scale), representing a 22.2% relative improvement over single-prompt LLM baselines (2.16/3.0) across all pedagogical criteria.
Download
Short Papers
Paper Nr: 5
Title:

From Generation to Certification: A Framework for Explainable and Taxonomy-Aware AI in Educational Assessment
Authors:

Antoun Yaacoub and Zainab Assaghir
Abstract: The rapid adoption of generative artificial intelligence (AI) in education has enabled scalable content creation and automated assessment design. However, existing approaches often focus on isolated aspects such as question generation, cognitive alignment, or feedback analysis, lacking an integrated framework that ensures pedagogical validity, transparency, and practical usability. This paper presents a unified framework for AI-driven educational assessment that combines taxonomy-aware generation, linguistic feedback adaptation, prompt-based control, and explainability mechanisms within a single system. Building upon the OneClickQuiz platform, we integrate Bloom’s and SOLO taxonomies to guide cognitive alignment, employ lightweight prompt engineering strategies to control generation, and incorporate linguistic analysis to adapt feedback across difficulty levels and pedagogical tones. To address trust and adoption challenges, we further introduce an explainability and certification layer that provides interpretable alignment evidence and audit-ready metadata for AI-generated questions. This enables educators to validate content quality and supports institutional requirements for transparency and accountability. We evaluate the proposed framework through a combination of automated analysis and user feedback, demonstrating improvements in cognitive alignment, controllability, and usability compared to baseline generative approaches. Our results highlight the importance of bridging generation, control, and explainability in AI-driven educational tools.
Download
Paper Nr: 11
Title:

AI-Enabled Automated Scaffolding for Undergraduate Students’ Learning to Debug and Reason
Authors:

Brian R. Belland, Wenpeng Yin, Zhuoyang Zou, Jack V. Mussoline and ChanMin Kim
Abstract: This paper is situated in a project in which we scaffolded 224 undergraduate students majoring in early childhood and elementary education at 2 universities to debug block-based programming when teaching with robots. Scaffolding is defined as cognitive and motivational support that allows learners to engage in higher-order thinking while engage in problem solving. In our earlier work, we found that when faced with bugs, such undergraduate students simply randomly deleted entire lines of code until the robot completed any action without stopping first. To help them learn to debug, we developed scaffolding grounded in the theory of abductive reasoning. In this paper, we describe how we used AI to dynamically adjust scaffolding support according to undergraduate students’ responses to scaffolding prompts. We fed the new AI-enabled automated scaffolding data from our prior studies in which preservice undergraduate students responded to scaffolding prompts. We report our evaluation of AI-enabled automated scaffolding in terms of its behavior in responding to text input. Results are discussed in light of the literatures on scaffolding automation and computer science education.
Download
Paper Nr: 12
Title:

Using Bias in Computer Vision AI to Teach Undergraduates Critical AI Literacy
Authors:

ChanMin Kim, Jungeun Ha and Brian R. Belland
Abstract: We introduce an innovative framework for critical AI education called Critical Eye for Ambient AI Aesthetics (CEAAA). We built the framework from two rounds of empirical studies among undergraduate students in the United States. In this paper, we report our second study findings along with the framework. The framework is grounded in critical race aesthetics and ambient identity cues. Critical race aesthetics guides individuals to notice bias in visual objects including ones produced by AI, and ambient identity cues help understand the impact of biased visual objects. Our studies were conducted in an undergraduate teacher preparation course on learning and instruction. The majority of participants were White. The campus is located in an area in which the majority of residents are Latinx. The instructional unit covered four weeks of the course (600 minutes in total). Data were collected from surveys, reflections, group discussions, class artifacts, and interviews. We analyzed data using a coding scheme based on the framework and applied theories. We found that participants demonstrated a high level of empathetic engagement and created bias-free materials for racial and gender minorities upon understanding societal impacts of biased AI on minoritized children. Findings and implications are discussed in connection to theories and the framework for critical AI education.
Download