Overview
Our project on AI Pedagogy investigates how teaching and learning strategies—long studied in psychology and education—can be translated into generative AI systems, especially large language models (LLMs). At the same time, it asks how these models can serve as experimental mirrors to refine our understanding of human learning. Rather than treating knowledge acquisition as passive data absorption, we explore dialogue, scaffolding, and exploration: the very processes through which humans learn best.
What we study
- Synthetic learning environments: We design structured, fictional ontologies—artificial but coherent “toy worlds”—to study learning without the bias of prior knowledge. These environments let us probe how LLMs acquire concepts from scratch.
- Teacher–learner dialogues with LLMs: Within these worlds, a teacher (an LLM with access to the ontology) interacts with a learner (an LLM or a human who starts naïve). Through structured exchanges, the teacher provides explanations, asks guiding questions, or responds to the learner’s curiosity. We test different strategies inspired by educational psychology: top–down and bottom–up exposition, learner-driven inquiry, teacher-guided questioning, and mixed forms of dialogue.
- Humans and LLMs, side by side: LLM simulations allow us to measure which strategies lead to faster and more robust knowledge acquisition. We then bring the same approaches to human experiments, to compare where LLMs behave like people and where they diverge. Such divergences highlight human-specific constraints (like working memory, inductive biases, or reliance on social cues), while convergences point to general principles of pedagogy that may hold across both humans and machines.
- Iterative discovery cycle: The project follows a cycle: formalise pedagogical strategies → test with LLMs → refine → test with humans → interpret differences → feed back into theory. This process allows us to move from broad pedagogical descriptions to mechanistic accounts of how knowledge is built through interaction. This translates into turning educational ideas into measurable AI protocols, testing them, comparing humans and machines, and using the insights to sharpen theories of how knowledge is built through interaction.
Why it matters
By embedding pedagogy into LLMs, we aim to produce agents that are not only more adaptive and transparent as learners or teachers, but also provide a scientific lens on human cognition. Generative AI becomes a kind of microscope for learning: making abstract theories of education testable, measurable, and comparable across humans and machines.
Methods at a glance
- Environment → Synthetic ontologies for structured, bias-free learning;
- Learner → Naïve LLM or human participant;
- Teacher → LLM providing explanations and guidance or human participant;
- Dialogue → Pedagogical strategies from psychology (scaffolding, inquiry, exposition);
- Evaluation → Efficiency of knowledge acquisition, quality of generalisation, and human–AI comparisons.
Project Members
Dr. Sabrina Patania, PhD
sabrina.patania@unimib.it
Dr. Luca Annese
luca.annese1@unimib.it
Dr. Cansu Koyuturk
c.koyutuerk@campus.unimib.it
Prof. Dimitri Ognibene
dimitri.ognibene@unimib.it
References
Patania, S., Annese, L., Koyuturk, C., Ruggeri, A., & Ognibene, D. (2025). AI Pedagogy: Dialogic Social Learning for Artificial Agents. Proceedings of the International Conference on Social Robotics (ICSR 2025).
Patania, S., Masiero, E., Brini, L., Piskovskyi, V., Ognibene, D., Donabauer, G., & Kruschwitz, U. (2024). Large Language Models as an active Bayesian filter: information acquisition and integration. In Proceedings of the 28th Workshop on the Semantics and Pragmatics of Dialogue, September, 11-12, 2024, University of Trento.