Continuously Adaptive Curricula: Using Learner Models to Restructure Instruction in Real Time

Abstract

Large language models can generate explanations and tutoring dialogue on demand, but most systems still rely on a curriculum that was fixed at the start of learning. The result is a path that does not change when a learner masters material early or repeatedly struggles with a prerequisite. This paper argues that the more important problem is not content generation but continuous reconstruction of the curriculum itself.We describe an architecture under development for the Mentrast platform that treats the curriculum as a living object. An AI tutor collects interaction evidence, a probabilistic learner model maintains estimates of mastery under uncertainty, and a curriculum engine uses those estimates to reorder, insert, or remove material. Drawing on work in intelligent tutoring systems and Bayesian knowledge tracing, we formalize curriculum selection as a repeated decision conditioned on the current learner state.A central practical constraint is also stated: if conversational evidence cannot be mapped reliably onto knowledge components, structural adaptation remains unreliable. The goal is a system that decides what the learner needs next rather than what was planned at the beginning.