Mapping teaching style spectra in LLMs : an initial contribution

UNAITE

Lucas JUNG, école Polytechnique Rina BADARIOTTI, Université Paris-Dauphine (PSL) Adam ZERAIKI, ESSEC et école Centrale Rémy SIAHAAN--GENSOLLEN Rémy, ENSAE

Training / education / pedagogy

A student has attended a full semester of lectures in operational research or political sociology of Europe and hasn't fully grasped the core concepts. Rather than passively re-reading notes, they turn to an LLM to actively revise: summarizing readings, unpacking lecture content, and testing their understanding through dialogue. Our case benchmarks available LLMs against each other and against the professor's explanations evaluating pedagogical clarity, reasoning style, and psychological support to identify which model best helps students learn without accumulating cognitive debt.

Today's LLMs often provide partial, superficial answers that create an illusion of understanding without its substance widening the gap between feeling informed and actually knowing. This use case challenges that pattern by asking not "can AI explain this?" but "can AI help a student think through this?" The goal is to identify which LLMs genuinely support active learning where the student does the work, not the AI. Beyond cognition, this raises a question of equity. Bourdieu's concept of cultural capital reminds us that academic success is shaped by inherited resources. A well-designed AI tutor could partially democratize access to high-quality, judgment-free academic support, offering struggling students what was previously only available through private tutoring or privileged networks. Finally, coupling pedagogy with psychological support could help isolated students re-engage. For many, academic difficulty is entangled with shame and loneliness. An AI that responds with both explanation and encouragement could lower the barrier to seeking help not replacing human connection, but making it easier to take the first step.

Faced with an ever-growing landscape of LLMs and the anarchic, informal benchmarks students spontaneously produce, our team wants to bring rigor where there is currently noise. We aim to objectively and quantitatively evaluate which available LLMs perform best for academic revision and reading synthesis going beyond surface-level output quality to analyze each model's reasoning style, user interface, and pedagogical depth. We go further by examining the psychological dimension of each model: how it responds to confusion, frustration, or repeated errors matters as much as whether its answers are correct. Ultimately, we hope to give students a concrete, evidence-based guide to using AI tools without accumulating cognitive debt, helping them think deeper, manage their time better, and remain the active agents of their own learning rather than passive consumers of generated content.