← back

Price

—

Travel

—

Access

—

Social

—

Latest developments in Large Language Models (LLMs) have elicited a divided response among educators, marked by enthusiasm for their potential as well as apprehension over their actual benefits in learning. Studies have shown that while many students are already using these models, nearly half of these interactions are student attempts to directly seek answers, undermining the productive struggle needed for learning and leading to potential deficits in cognitive development (famously dubbed “the cognitive debt of AI”). In this talk, I will present my group's vision towards addressing this grand challenge by training language models that incorporate pedagogical theories from learning science and are designed to teach (and not just provide answers), spot student mistakes, understand their deeper misconceptions and offer targeted guidance, and track students’ knowledge evolution and adapt to their needs. I will demonstrate this vision by describing our approach to develop a pedagogically aligned conversational math tutor building on a series of innovations: (a) new ways to collect tutoring data at scale, (b) new benchmarks for evaluating models on key pedagogical criteria, and (c) new ways of adapting language models to pedagogy using supervised finetuning, offline and online reinforcement learning, and simulation-based training. Then, moving beyond problem-level assistance, I will present new techniques that bridge classical models of human learning such as knowledge tracing with language models, allowing them to account for a student’s learning progress and provide personalized assistance, e.g. by recommending what to learn next. Finally, I will conclude by describing the translational impact of our work in both K-12 and university contexts, and past and ongoing studies with human learners.

• Other