Dissertation Title: Learning to Teach: Models for Semantic and Adaptive Personalization of AI Tutors

Abstract: 

AI tutoring systems have the potential to make education affordable, equitable, and effective. However, despite recent advances in large language models (LLMs), mainstream use of AI as tutors remains limited due to a lack of pedagogically grounded frameworks. Previous work has shown that effective tutoring involves not only linguistic fluency, but also the ability to (i) to effectively engage in open-domain dialog to build rapport and (ii) adapt dynamically to feedback cues from a learner.

This thesis addresses these limitations by providing two key contributions. First, we study the cognitive and psychological processes humans use when engaging in conversations. We propose a novel probabilistic approach using Markov Random Fields (MRF) to augment existing closed-source and open-source LLMs (GPT-4o, Gemini and LLaMa) for improved next-utterance generation. Using human evaluations, we show that our augmentation approach significantly improves the performance of existing state-of-the-art LLMs for open-domain conversational agents.

Next, we explore the challenge of dynamic pedagogical adaptation. Learners provide feedback in different forms (facial expressions, gaze, questions, seeking help, etc.), and the interpretation of this feedback may differ from learner to learner. However, existing approaches like reward shaping, policy shaping and preference-based RL assume a static interpretation of feedback labels. To this end, we present an adaptive reinforcement learning algorithm that generalizes to multiple labels and autonomously interprets the meaning of feedback cues during online deployment.

Through data from human subjects, we show that the algorithm can learn from feedback cues and outperform any static interpretation in a simulated environment. We provide rigorous studies showing that adaptive RL is invariant to noise and bias in feedback cues as well as the choice of MDP. We further outline the conditions for employing such an algorithm. Finally, we integrate our adaptive RL algorithm with an existing AI tutor for Python programming to enable it to adapt to each user.  We study the characteristics of our algorithm against baseline RL in real-world tutoring systems and present our findings and recommendations.

By integrating cognitive modeling and adaptive RL into AI tutors, this thesis contributes to the design of intelligent, personalized, and pedagogically grounded educational systems.

Cynthia Breazeal, Professor of Media Arts and Sciences, MIT; MIT Dean for Digital Learning
Alex P. Pentland, Toshiba Professor of Media, Arts & Sciences, MIT
Hal Abelson, Professor of Professor of Computer Science and Engineering, EECS, MIT
Hae Won Park, Research Scientist, MIT Media Lab, MIT