Illuminate: Adaptive AI Tutoring for Research and Practice
AI tutoring that adapts to what you know: switching from patient instructor to challenging debate partner as your mastery grows.
Duration: January 2026 – December 2027
Status: Ongoing
Educational Level: Tertiary Level
Topic: Artificial Intelligence AI, Data Science for Education
Keywords: AI Tutoring, Bayesian Knowledge Tracing, Adaptive Learning, Personalized Education, Learning Analytics
Initial Situation
Current AI tutoring systems treat all students the same way, regardless of what they already know. This ignores a basic principle of learning: the best way to teach someone depends on their starting point. Teaching methods that work well for beginners often fail for advanced learners, and approaches that challenge experts can overwhelm newcomers. Step-by-step guidance helps those just starting out but bores those who have moved beyond the basics; open-ended exploration stimulates advanced learners but frustrates beginners. This one-size-fits-all approach also raises concerns about students becoming too dependent on AI. When systems simply hand over answers, students miss out on the struggle that builds real understanding; they receive solutions without developing the ability to solve problems themselves. Research in this field faces its own problem: fragmentation. Different research teams build their own separate systems using different approaches, making it nearly impossible to compare findings or build on each other’s work. Good ideas stay trapped in individual labs instead of advancing the field as a whole. These issues matter for everyone involved: students get help that does not match their needs, researchers cannot systematically test what actually works, and teachers lack clear guidance on how to use AI effectively in their courses.
Objectives
Illuminate builds an adaptive AI tutoring system that adjusts its teaching based on what each student knows. The system tracks learner progress and switches between teaching styles: from patient step-by-step instruction for beginners to challenging discussion for advanced learners. Beyond this, Illuminate provides a shared research platform where different tutoring approaches can be tested and compared under controlled conditions. This enables researchers to discover which strategies work best for which learners, advancing both the science and practice of AI-assisted education.
Method
We use Bayesian statistical methods to model student knowledge, capturing both current understanding and uncertainty in our estimates.This approach supports four integrated components: a knowledge tracking model estimating what each student knows, an exercise selection system choosing appropriate problems, a set of tutoring strategies from direct explanation to guided questioning, and a selection mechanism matching approaches to learner states. The system enables adaptive experimentation: randomized controlled designs compare teaching strategies across knowledge levels, directly testing predictions from learning science, particularly the expertise reversal effect. The pilot in a mathematics course at the University of Bern provides the testbed for these investigations.
Planned Translation
Illuminate will be piloted in an introductory mathematics course at the University of Bern under real teaching conditions. This course already uses a digital learning environment with formative assessments and exercises linked to specific learning goals, providing an ideal foundation for integration. The system is designed for modular deployment, allowing iterative refinement based on feedback from students and instructors. Findings will directly inform the development of adaptive tutoring features. For educators, the platform provides insight into student learning paths, enabling earlier intervention. For institutions, integrated data collection supports evidence-based decisions about teaching effectiveness.
For research, Illuminate provides infrastructure enabling systematic comparison of AI tutoring approaches. By allowing researchers to test strategies within the same platform, the project supports reproducible findings and will accelerate progress toward evidence-based AI tutoring. For practice, the system offers students personalized support that adapts to their knowledge level, potentially improving learning outcomes. Educators gain real-time insight into student progress, enabling earlier intervention. Impact measurement includes: comparing learning outcomes between groups through course assessments and counting research outputs such as publications and theses produced using the platform.
