Beyond Expert-Annotated Labels: An Adaptive Label Learning Method for Knowledge Tracing

ICLR 2026 Conference Submission15876 Authors

Keywords: Knowledge Tracing, AI for Education, Assessment, Adaptive Label Learning

TL;DR: We propose an adaptive framework for optimizing question groupings in Knowledge Tracing models, which significantly enhances prediction accuracy and uncovers hidden cognitive mechanisms among questions.

Abstract: Knowledge Tracing (KT) serves as an indispensable technology in intelligent tutoring systems (ITS), aiming to predict learners' future performance based on their past interactions. Current KT models commonly use predefined knowledge concept (KC) labels to improve prediction accuracy. These labels provide grouping information about questions, allowing models to infer learners' performance on low-frequency questions. However, the subjectivity of human labeling may not accurately reflect which questions share similar cognitive processes, potentially limiting the models' performance. To address this, we redefine KT as a problem of learning from question groupings and introduce an adaptive framework that iteratively refines groupings through alternating optimization. We initiate with random groupings and freeze them to optimize the KT model with gradient descent, then select the loss-minimizing configuration by computing the loss for each possible reassignment of questions to different groups under continuous assignment probabilities, repeating this process until convergence. We evaluate our approach on real-world ITS datasets, incorporating the optimized groupings into different KT models instead of KCs, which markedly improves model performance and achieves state-of-the-art results. Further experiments uncover the underlying semantic connections between our automatic groupings and prior KCs, revealing potential similarities in cognitive mechanisms among KCs, providing new insights and research directions for educational and cognitive sciences. Code is available in the supplementary materials.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 15876