Effective Personalized AI Tutors via LLM-Guided Reinforcement Learning

Generative AI is reshaping work faster than education and training systems can adapt, creating an urgent need for efficient skill acquisition. Existing adaptive learning algorithms have limited visibility into learners’ progress, constraining their ability to personalize training effectively. We address this limitation by designing an AI-powered learning platform that integrates large language models (LLMs) with reinforcement learning (RL): the LLM generates exercises, provides tutoring support, and extracts rich behavioral signals from learner interactions, while the RL agent uses these signals to adaptively sequence practice question difficulty to maintain productive struggle matched to each learner’s evolving skill level. In partnership with the Taipei City Government and the American Institute in Taiwan, we launched a five-month “AI for Python Learning” program to evaluate this platform by conducting a randomized controlled trial (RCT) with 1,047 high school students. Despite a strong control condition that all students had access to an LLM tutor, the treatment group receiving additional RL-driven adaptive sequencing achieved a 0.15 standard deviation increase in in-person, paper-based exam performance—equivalent to 6–8 months of additional schooling. Mediation analysis suggests that these gains were driven by increased learner engagement. These results provide large-scale field evidence that combining rich interaction signals via LLM with adaptive learning algorithm can significantly accelerate skill acquisition beyond access to generative AI alone.