Timely and accurate identification of student misconceptions is key to improving learning outcomes and pre-empting the compounding of student errors. However, this task is highly dependent on the effort and intuition of the teacher. In this work, we present a novel approach for detecting misconceptions from student-tutor dialogues using large language models (LLMs). First, we use a fine-tuned LLM to generate plausible misconceptions, and then retrieve the most promising candidates among these using embedding similarity with the input dialogue. These candidates are then assessed and re-ranked by another fine-tuned LLM to improve misconception relevance. Empirically, we evaluate our system on real dialogues from an educational tutoring platform. We consider multiple base LLM models including LLaMA, Qwen and Claude on zero-shot and fine-tuned settings. We find that our approach improves predictive performance over baseline models and that fine-tuning improves both generated misconception quality and can outperform larger closed-source models. Finally, we conduct ablation studies to both validate the importance of our generation and reranking steps on misconception generation quality.

In mathematics tutoring, using appropriate instructional discursive strategies, called "talk moves'', is critical to support student learning. Training tutors in the appropriate use of talk moves is a key component of tutor development programs. However, tutor development at scale is a challenge. Recent research has shown that automatic talk moves classification of tutorial discourse can facilitate large-scale delivery of personalized talk moves feedback. In this paper, we build on this work and share our current progress using large language models to classify talk moves in transcripts of tutoring sessions. We report classification results from fine-tuned models, prompt optimization, and supervised embedding vectors classification. The fine-tuned strategy performed best, yielding better performance (.87 macro and .93 weighted f1 score in predicting expert labels) than the current state-of-the-art RoBERTa model. We discuss trade-offs across methods and models.

The utility of assessment systems lies in their capacity to transform observations of student behavior into meaningful inferences about learning, knowledge, and skills. Common practice is to use latent variable models and produce scores on scales. However the simplicity of these psychometric models may filter out potentially valuable information present in student behavior. In particular, scale scores are not optimized to support granular instructional decisions. Machine learning offers promising alternatives, but proposed deep learning architectures are not ideally suited for operational testing conditions involving sparse data and shifting category labels for test questions.