System-One large language model improves knowledge tracing with few learners
Can a System-One LLM Perform Knowledge Tracing When Few or No Learners Are Logged?
Computation and LanguageComputers and SocietyMachine Learning
Summary
Knowledge tracing helps track what learners know, but it usually needs lots of data to work well. The authors tested a type of large language model (LLM) called System-One that can predict student knowledge quickly without needing much data. On seven different learning datasets, this model did better than many deep learning models even with very few logged learners. Adding some data from similar learners helped it improve further, but for new questions where all learners are known, traditional methods still work better.
What this means in practice
- •For online course developers: Build knowledge tracing models for new courses quickly without needing large learner datasets.
- •For educational platform engineers: Integrate fast and cost-effective learner knowledge prediction using off-the-shelf LLMs when user data is limited.
Authors
Unggi Lee, Haeun Park
Abstract
Knowledge tracing (KT) models need many logged learners, so a new course or platform starts without a usable model. In LLM-based KT the LLM generates the answer, which we call System-Two; it is either fine-tuned on the target data or reasons and votes over ten samples, which is slow and gives coarse probabilities. We ask whether an off-the-shelf System-One LLM, which returns a probability for a typed question directly in a single pass, can perform KT when few or no learners are logged. On seven datasets, Jev without any data from the target platform reaches a mean AUC of .706, above the best of 28 deep KT models trained on 8 learners (.689) and above System-Two Thinking-KT on all seven datasets (.650) at about 1/100 of its API cost. Adding examples and a similar-learner statistic from the logged learners (JevKT) raises this to .722; JevKT stays significantly ahead of deep KT up to 16 learners and ahead on average up to 64, and supervised KT catches up between 64 and 128 learners. Among the readers we tested, the gain is specific to Jev, since three other LLMs queried with the byte-identical typed request through the official System-One adapter fall below it on all seven datasets, and reader swaps and contamination checks find no evidence that the input format or memorised data explain the gain. For new learners the advantage holds from their first interactions, whereas on unseen items with all learners logged, deep KT remains ahead.