Pedagogical AI Agents

Educational chatbots built on large language models show real promise for learning outcomes, but most delegate pedagogical decisions such as what to teach and how to the model itself, making their tutoring strategies difficult to trace, evaluate, or reproduce, and leaving them prone to the error rates and domain variability that still make LLMs unreliable as autonomous instructors. This work instead formalizes the instructor’s own pedagogical reasoning into an explicit, inspectable layer: identifying what a student needs, selecting a didactic approach, and only then handing the wording over to a language model acting purely as a linguistic executor. The longer-term ambition is a Socratic companion that accompanies students throughout a course rather than just answering in isolation, treating an unclear question as an opportunity to guide the student toward a better one rather than as a failure to route, and keeping each interaction anchored in the instructor’s own teaching material and practice.

A first system, deployed for a university course on dimensional modeling, was evaluated on real student questions collected in production.

A second, more ambitious system extends this architecture with a library of discursive strategies and a multi-zone learner model, moving from a simple assistant towards a full Intelligent Tutoring System.