From Language Models
to Learning Companions:
Avatars and GenAI in CS Education 

This two-part workshop, organized by CoDiS Lab Graz, brings together researchers and educators to explore the design and implementation of avatar-based learning support for computational thinking. The program combines short presentations, collaborative design sessions, and hands-on prototyping activities. Participants will have the opportunity to share their own work, engage in peer exchange, and experiment with tools and methods for creating interactive learning agents.

The workshop is free of charge for all invited participants. Coffee and snacks will be provided during the sessions; participants are kindly asked to cover their own travel and accommodation expenses. Please bring a laptop for the prototyping activities. 

Location

On-site: Institute of Human-Centred Computing, Graz University of Technology, Sandgasse 36/I, 8010 Graz - Showroom (DHEG136E)

Virtual (Webex): https://tugraz.webex.com/tugraz/j.php?MTID=m54b1bb5a98a94ec4c3b11449dc32cb00

Program

Monday, July 28th, 2025

12:15 - 13:30 | Lunch (Napo at TU Graz)

13:30 - 14:00 | Introduction

14:00 - 15:30 | Presentations

The development of Sourcerer: An Intuitive Interface for Explainable Scientific Document Summarization 

Anton Schleinitz

NeutrinoReview - LLM supported Literature Search

Elias Sandner

Do LLMs Change Students' Code Style? A Static Code Analysis of CS1 Submissions Before and After LLM Proliferation

David Kerschbaumer

15:30 - 16:00 | Refreshment Break

16:00 - 17:30 | Hands-on Session

Avatar Lab I: Designing Task-Driven Agents for Computational Thinking 

Slides | Miro Board

17:30 | Social Event/Dinner

Tuesday, July 29th, 2025

09:00 - 09:30 | Welcome and Recap

09:30 - 10:30 | Presentations

How to Design and Deliver Courses for Higher Education in the AI Era

Muhammad Imran Taj

Automated Feedback on Student-Generated UML and ER Diagrams Using Large Language Models

Sebastian Gürtl

10:30 - 11:00 | Refreshment Break

11:00 - 12:00 | Presentations

AI-Based Pedagogical Agents for Learning Computational Thinking: Learners’ Perceptions and Preferences

Julia Pöschko

Designing and Instructing an AI-Based Peer Tutor for Learning Computational Thinking in Higher Education

Saba Soleimani

12:00 - 13:00 | Lunch

13:00 - 15:00 | Hands-on Session

Avatar Lab II: Implementing Agent Behavior for Learner Interaction

15:00 - 15:30 | Conclusion & Farewell

Materials

Feel free to upload your slides or any additional materials to this folder: https://cloud.tugraz.at/index.php/s/aBNzfR3dT3zSnoX

ABSTRACTS

The development of Sourcerer: An Intuitive Interface for Explainable Scientific Document Summarization

Anton Schleinitz (Institute of Human-Centred Computing - Graz University of Technology)

In 2024, the Institute of Human-Centred Computing (formerly known as Institute of Interactive Systems and Data Science) at Graz University of Technology, in collaboration with CERN in Geneva, Switzerland, introduced a system for generating explainable, accurate, and trustworthy summaries of long scientific documents. To make this system more accessible, an intuitive user interface, named Sourcerer, was developed as part of this thesis. Sourcerer enables users to interact with the summarization system and engage directly with scientific documents through a chat interface powered by locally running large language models. The summarization system provides candidate source passages for each summary sentence, supporting transparency and traceability. A core requirement of the platform was to allow users to visually explore summaries alongside their source candidates, enhancing the system’s explainability and usability.  

Beyond summarization, the application allows users to store uploaded documents for future reference and to extend its capabilities by adding other  large language models. These models can be integrated locally by specifying their Hugging Face repository and model name.  

The entire system was developed with simplicity in mind, aiming to make powerful language technologies accessible through a clean and user-friendly interface.

NeutrinoReview - LLM supported Literature Search

Elias Sandner (CERN; Institute of Human-Centred Computing - Graz University of Technology)

NeutrinoReview is being developed as a joint venture between the WHO, CERN, and the CoDiS Lab at TU Graz. Designed to support researchers in conducting systematic reviews, the current version of the tool reduces human workload by automating several tasks, including the retrieval of bibliographic data, deduplication, and LLM-based literature filtration based on flexible eligibility criteria.

While NeutrinoReview is specifically tailored to meet the rigorous requirements of evidence synthesis in the medical domain, it can also support evidence-based research in other fields and serve as a search engine for related work.

The presentation will provide an overview of the project, a live demonstration of the tool, and suggestions for how it could be used by computer science researchers and students.

Do LLMs Change Students' Code Style? A Static Code Analysis of CS1 Submissions Before and After LLM Proliferation

David Kerschbaumer (Institute of Human-Centred Computing - Graz University of Technology)

The rise of large language models (LLMs), such as ChatGPT, has a massive impact on computer science education, yet their influence on fundamental coding practices remains unclear. In this project, we investigate how the use of LLMs affects the code style of first-semester computer science students. We compare CS1 submissions before and after the rise of LLMs, maintaining consistent course setup and assignments. 

Using static code analysis, we examine metrics such as comment density, C library usage, and code complexity. Machine learning classifies 2024 submissions based on the likelihood of AI assistance. Our findings reveal considerable changes in student coding behavior following the introduction of AI tools. This work offers valuable insights for educators in computer science on the practical application of AI technologies by students in a CS1 course.

How to Design and Deliver Courses for Higher Education in the AI Era

Muhammad Imran Taj (College of Interdisciplinary Studies - Zayed University)

Technological breakthroughs in Generative Artificial Intelligence (AI) are challenging education. We argue that higher education will only cope with the era of AI when we reduce the reliance on textbooks and memorization, and take deliberate steps to integrate AI into the design and delivery of courses and exams. This work presents some strategies for this, based on Delors report on education. We demonstrate how we used these strategies in multiple courses in cybersecurity, programming, English language teaching, and art. As for course delivery, we automated the Socratic teaching approach by simulating a chatbot called the AI Socrates Chatbot. We employed this chatbot in our classes to offer personalized learning experiences for our students. To evaluate the proposed instructional design, we analyzed data from seven exams in which our students were allowed to use ChatGPT during the exam. The findings indicate that students’ performance was not correlated with the level and way they used ChatGPT to answer the exam questions.

Automated Feedback on Student-Generated UML and ER Diagrams Using Large Language Models

Sebastian Gürtl (Institute of Human-Centred Computing - Graz University of Technology)

UML and ER diagrams are foundational in computer science education but come with challenges for learners due to the need for abstract thinking, contextual understanding, and mastery of both syntax and semantics. These complexities are difficult to address through traditional teaching methods, which often struggle to provide scalable, personalized feedback, especially in large classes. We introduce DUET (Diagrammatic UML & ER Tutor), a prototype of an LLM-based tool, which converts a reference diagram and a student-submitted diagram into a textual representation and provides structured feedback based on the differences. It uses a multi-stage LLM pipeline to compare diagrams and generate reflective feedback. Furthermore, the tool enables analytical insights for educators, aiming to foster self-directed learning and inform instructional strategies. We evaluated DUET through semi-structured interviews with six participants, including two educators and four teaching assistants. They identified strengths such as accessibility, scalability, and learning support alongside limitations, including reliability and potential misuse. Participants also suggested potential improvements, such as bulk upload functionality and interactive clarification features. DUET presents a promising direction for integrating LLMs into modeling education and offers a foundation for future classroom integration and empirical evaluation.

AI-Based Pedagogical Agents for Learning Computational Thinking: Learners’ Perceptions and Preferences

Julia Pöschko (Educational Psychology - University of Graz)

Pedagogical agents are virtual characters in computerized learning environments that support student  learning. While pedagogical agents have been present for decades, recent technological advancements in generative AI revolutionize their use. AI-based pedagogical agents provide a range of benefits for learning, such as scalability, personalization, and on-demand use. However, to design agents in a way that enables successful learning, we need to consider learners’ perspectives on AI-based pedagogical agents. Therefore, we pose the following research question: What are learners’ perceptions and preferences regarding learning with an AI-based pedagogical agent? As learning content, we chose computational thinking. 

We plan to conduct a within-subjects vignette experiment with university students from different disciplines. Students will be presented with short video vignettes that give an impression of learning computational thinking with an AI-based pedagogical agent in the role of a peer. The vignettes vary in four factors: 1) modality of interaction, 2) visual representation of the agent (avatar), 3) avatar gender, and 4) avatar anthropomorphism. Subsequently, we will determine students’ preferred agent as well as differences  between agents according to students’ ratings of learning-related variables, such as trust, social presence, expected learning effectiveness, and willingness to use. Potential covariates such as students’ gender, academic self-concept, personality, and stereotype beliefs will be assessed additionally for a more nuanced picture of students’ preferences.

The findings of this study will inform the design of AI-based pedagogical agents based on learners’  judgements. This is a first step toward a future of learning that meaningfully utilizes AI through integration with pedagogical agents.

Designing and Instructing an AI-Based Peer Tutor for Learning Computational Thinking in Higher Education

Saba Soleimani (IT:U - Interdisciplinary Transformation University, Linz, Austria)

This study investigates the design and implementation of an AI-based pedagogical agent that adopts the role of a peer tutor to support development of computational thinking (CT) skills in higher education. Focusing specifically on decomposition skill, the research examines how structured, role-switching interactions between learners and the AI agent affect learning outcomes, collaboration patterns, and learner perceptions. Drawing on established instructional strategies and grounded in the 3P model (Presage, Process, Product), the study integrates both cognitive and socio-emotional design features in the agent’s behaviour to scaffold learning effectively. The experimental design involves multiple phases in which learners alternate between tutoring and being tutored by the agent, with performance and perception data collected at each stage. Key variables under investigation include learner satisfaction, self-concept, social presence, trust, and the moderating role of prior knowledge. By combining quantitative and qualitative data from interaction logs, performance tests, and perception surveys, the study aims to provide an understanding of how AI-based peer tutoring can enhance computational reasoning and support personalised, engaging, and equitable learning experiences. The findings will inform the design of socially responsive pedagogical agents and contribute to the growing body of interdisciplinary research at the intersection of AI, education, and learning sciences.