Problems With Large Language Models for Learner Modelling: Why LLMs Alone Fall Short for Responsible Tutoring in K--12 Education

TL;DR

The study investigates whether general-purpose LLMs can serve as responsible tutors in K--12 education by comparing them to Deep Knowledge Tracing (DKT) on the ASSISTments dataset. Despite fine-tuning, LLMs lag behind DKT in next-step prediction ($AUC$), temporal coherence, and stable mastery trajectories, requiring substantially more compute. The results argue that LLMs alone cannot reliably track evolving learner knowledge and should be integrated with predictive learner models in hybrid, neural-symbolic frameworks to align with responsible AI principles. Practically, this supports a hybrid tutoring paradigm where knowledge-tracing signals constrain LLM outputs to enable accurate adaptivity while leveraging LLMs for richer feedback and explanations.

Abstract

The rapid rise of large language model (LLM)-based tutors in K--12 education has fostered a misconception that generative models can replace traditional learner modelling for adaptive instruction. This is especially problematic in K--12 settings, which the EU AI Act classifies as high-risk domain requiring responsible design. Motivated by these concerns, this study synthesises evidence on limitations of LLM-based tutors and empirically investigates one critical issue: the accuracy, reliability, and temporal coherence of assessing learners' evolving knowledge over time. We compare a deep knowledge tracing (DKT) model with a widely used LLM, evaluated zero-shot and fine-tuned, using a large open-access dataset. Results show that DKT achieves the highest discrimination performance (AUC = 0.83) on next-step correctness prediction and consistently outperforms the LLM across settings. Although fine-tuning improves the LLM's AUC by approximately 8\% over the zero-shot baseline, it remains 6\% below DKT and produces higher early-sequence errors, where incorrect predictions are most harmful for adaptive support. Temporal analyses further reveal that DKT maintains stable, directionally correct mastery updates, whereas LLM variants exhibit substantial temporal weaknesses, including inconsistent and wrong-direction updates. These limitations persist despite the fine-tuned LLM requiring nearly 198 hours of high-compute training, far exceeding the computational demands of DKT. Our qualitative analysis of multi-skill mastery estimation further shows that, even after fine-tuning, the LLM produced inconsistent mastery trajectories, while DKT maintained smooth and coherent updates. Overall, the findings suggest that LLMs alone are unlikely to match the effectiveness of established intelligent tutoring systems, and that responsible tutoring requires hybrid frameworks that incorporate learner modelling.

Figures (6)

• Figure 1: Potential sources of bias during the initial development of LLMs, adapted from Lee et al. lee2024life.
• Figure 2: Illustration of the evolving paradigms in artificial intelligence, adapted from Dhar dhar2024paradigm.
• Figure 3: Dataset characteristics after pre-processing: (a) distribution of interactions per student, (b) top 20 most frequent quizzes, and (c) top 20 most frequent skills.
• Figure 4: ROC curves comparing prediction performance across all models.
• Figure 5: Heatmaps of predicted skill-mastery probabilities for the test-set student with ID 81439: (a) DKT model, and (b) fine-tuned LLM model. Annotated cell values highlight points where the model’s mastery update was incorrect, while the white reference line represents the correct, temporally stable update trajectory for each skill—indicating how mastery should rise or fall based on the learner’s previous state and most recent response.