Context
The integration of computer programming into secondary education has become a strategic priority for modern societies, as digital skills are increasingly essential for economic participation and social inclusion. However, learning to program remains cognitively demanding, and teachers often lack the time and resources to provide individualized feedback at scale. Such support can be delivered through adaptive feedback, that is, personalized information communicated to the learner with the aim of modifying their thinking or behavior to improve learning. Recent advances in Artificial Intelligence, particularly Large Language Models (LLMs), open new possibilities for automating personalized support through adaptive feedback. Yet, existing systems struggle to control the pedagogical quality of AI-generated feedback [4], the key challenge lies in providing sufficient help to prevent discouragement while avoiding excessive assistance that may hinder learning.

Hypothesis and research questions
This thesis builds on the hypothesis that the accuracy and pedagogical appropriateness of AI-generated feedback can be improved by coupling LLMs with predictive models based on Graph Neural Networks (GNNs) [5], trained on large-scale historical interaction data, which inform a pedagogical framework aimed at regulating feedback generation.
This research addresses four main questions :
(1) How can tabular and temporally structured learner historical interaction traces be modeled as interaction graphs in order to accurately predict performance, behavior, and pedagogically-related indicators ?
(2) Which Graph Neural Network architectures and training strategies are most effective for predicting indicators from these interaction graphs ?
(3) How can these predictive indicators be integrated into LLM-based systems through prompt engineering and multi-agent orchestration to generate feedback that is both adaptive and pedagogically controlled ?
(4) What is the impact of the proposed system on the pedagogical quality of generated feedback, learning outcomes, and learners’ and teachers’ acceptance ?

Work plan
The proposed system will be implemented in the Pyrates application , developed and maintained by Branthôme [1–3]. Pyrates is a serious game designed to introduce fundamental programming concepts through the control of a pirate character in a platform game using Python code. This application has been widely adopted in French secondary education, with over 350,000 games played, which has generated an extensive dataset of historical interaction traces.
The thesis work will be composed of several phases :
1/ The pedagogical framework will be established by educational sciences experts, and the predictive indicators most relevant to inform the framework will be defined.
2/ Pyrates’ historical interaction traces (relational data) will be structured as property graphs, and various representations will be explored to better match indicator prediction. Multiple GNN models will be trained and compared to identify the most effective architecture for each indicator.
3/ The pedagogical framework will be implemented via LLM prompt engineering [6]. The orchestration of multiple LLM agents [7] will be investigated, and the Pyrates interface will be adapted to integrate interactions with these agents.
4/ The solution will be evaluated in the field through experiments conducted in French secondary schools.

Candidate Profile
We are seeking highly motivated PhD candidates with a strong background in Computer Science (MSc), and demonstrated excellence in Machine Learning and Data Science. Applicants should have solid programming skills in Python and Web technologies. Strong written and oral communication skills in English are essential.
This PhD position is for a duration of three years and will ideally be based in Lannion (on the northern coast of Brittany, France). The PhD is expected to start in October 2026.
Interested candidates should first contact Matthieu BRANTHÔME by email (matthieu.branthome@irisa.fr) and provide the following documents:
- a detailed curriculum vitae,
- academic transcripts (to assess the required skills),
- either letters of recommendation or contact details of referees who can support the application.

More informations : https://www.irisa.fr/phd-subject/2026-01/enhancing-ai-generated-adaptive...