LLMs Underutilized in Tutoring, Study Finds
A new preprint explores the use of large language models in foreign language education, finding they can be effective tutors. However, the authors conclude that most current deployments merely automate conventional teaching methods. The study argues that these systems fail to leverage the full potential of AI for real-time adaptation, knowledge tracing, or reinforcement learning-based content sequencing.
- To create adaptive learning experiences, a contextual multi-armed bandit framework can be used to select personalized learning actions. In this model, a student's prior knowledge state serves as the "context," and the various learning activities are the "arms" of the bandit, with the goal of maximizing future assessment performance. - Knowledge Tracing (KT) models a student's understanding of concepts over time by analyzing their interactions with learning materials. Deep learning-based approaches like Deep Knowledge Tracing (DKT) have become state-of-the-art for predicting student performance. - Reinforcement learning (RL) can optimize the sequence of learning materials by treating it as a policy to be learned. The system can adjust the difficulty of questions in real-time to maximize a student's learning outcomes based on their previous answers. - For early reading tutors, speech recognition systems must be specifically trained on children's voices, which have different pitches and patterns than adults, and be robust to noisy classroom environments. Companies like SoapBox Labs and Amira Learning specialize in this technology for educational applications. - A 2021 study at Brewbaker Primary School found that second-grade students using the AI-powered reading tutor Amira doubled the number of words they could correctly read per minute after two months. Another study showed that Amira improved the vocabulary of third-grade English language learners more than one-on-one human tutoring. - AI can enhance phonics instruction by providing personalized, interactive exercises that give immediate feedback on pronunciation. Tools like Magic School AI can even generate decodable texts tailored to specific phonics patterns a student is learning. - Effective AI tutors for reading are grounded in the Science of Reading, focusing on five key pillars: phonemic awareness, phonics, fluency, vocabulary, and comprehension. - Federated learning is an emerging approach that allows AI models to learn from user interactions across different devices without centralizing sensitive student data, which enhances privacy.
