LLM4Edu

Learning Large Language Models on Graphs for Education

Educational platforms generate rich interaction data, but adapting large language models to real educational settings is still expensive, opaque, and often hard to justify to stakeholders. LLM4Edu develops graph-based methods to build education-specific benchmarks and to fine-tune LLMs efficiently and transparently, in collaboration with Macmillan Learning and Holtzbrinck Publishing Group.

What we do

We pursue three tightly connected research directions:

  1. Data and benchmarks
    We turn educational interaction data into structured, graph-based representations (learners, content, activities, dependencies). Based on this, we create a benchmark dataset that enables systematic evaluation of personalization in educational scenarios.

  2. Efficient and interpretable adaptation
    We develop parameter-efficient fine-tuning methods that keep the core LLM largely frozen and train a lightweight Graph Neural Network (GNN) for adaptation. The goal is not just to reduce training cost but also to make the adaptation mechanism more interpretable than standard PEFT pipelines. Parameter-efficient fine-tuning is increasingly recognized as essential for adapting large models to specific domains under practical resource constraints.

  3. Evaluation in realistic educational workflows
    We benchmark LLMs and graph-based variants, analyze what matters for educational tasks, and validate the approach in settings that reflect real learning and teaching processes.

Why it matters

AI can enhance personalized learning by adapting content and pacing to individual needs, improving engagement and outcomes beyond traditional one-size-fits-all approaches. However, the technical challenges of adapting large models to real educational data limit adoption in practice. LLM4Edu targets this bottleneck by combining efficiency and transparency:

  • Scalable personalization: By reducing the need for full re-training, our methods make it feasible to tailor models to diverse curricula and learner profiles without prohibitive compute costs.
  • Transparent adaptation: Educators and practitioners need to understand why a model behaves as it does. A graph perspective helps reveal what structures in the data drive model adaptation.
  • Benchmarking rooted in educational reality: Standard NLP benchmarks do not reflect the structural dependencies of learning data. Our dataset enables evaluation that aligns with real teaching and learning needs.
  • Responsible deployment: Clear indicators of adaptation and systematic evaluation support more reliable and accountable AI in education.

Get involved

Students: Thesis / HiWi opportunities are available; please contact Shuzhou Yuan (project lead) or see /student_thesis/.
Media: For expert interviews on AI in education, LLMs, and trustworthy AI, see /media/.

Lead: Shuzhou Yuan
Academic Advisor: Michael Färber
Partners: Macmillan Learning; Holtzbrinck Publishing Group
Start: 01/2025 · Duration: 17 months

References