Adaptive LLM Tutoring Boosts Student Engagement and Efficiency

Large Language Models (LLMs) hold significant promise for personalizing education, but current static-prompt tutoring systems often struggle to adapt effectively across diverse academic subjects. This limitation can hinder their ability to cater to individual student needs and maintain engagement. To address this, researchers developed and tested an innovative adaptive tutoring system. This system incorporates subject-aware prompting, which dynamically adjusts its approach based on 14 pedagogical features extracted from student-tutor interactions, such as tutor scaffolding and student understanding. A prompt routing model, initially trained in a simulation environment, was then deployed for online adaptation with actual high-school students. The simulation benchmark demonstrated that the router significantly outperformed two static baseline systems. Subsequent A/B testing with 359 students (656 conversations) confirmed the model's ability to transfer its adaptive strategies from simulation to real-world scenarios, switching between analytical and scaffolding learning approaches. This adaptive prompt selection mechanism not only improved instructional efficiency but also maintained pedagogical quality and reduced the number of interactions by approximately three turns. While a greedy router achieved a comparable exercise conversion rate, a stochastic router that sampled strategies led to an even higher conversion rate, highlighting the benefits of strategic adaptability.