UK study of 119K students finds math ability is one dominant factor

A new study published on arXiv (2607.26063) challenges assumptions in personalized learning by analyzing whether mathematical ability consists of discrete, sequentially acquired skills or a single overall factor. The team—Benjamin Mawdsley, Tom Quilter, Richard Turner, Sarah Jackson, and Paul Edwards—used a Bernoulli Mixture Model to cluster pass/fail results from 119,034 students who took 13 national-level exams in the United Kingdom. The dataset was collected by an unnamed platform, providing a national-scale test of machine learning in education. The model searched for latent populations indicative of discrete skill-sets, but found very few distinct clusters. Instead, overall student ability was the dominant factor, confirmed by high linear correlation between cluster probability distributions.

The best-performing clustering model achieved 78% accuracy—competitive with more complex models in the literature while being more explainable. When compared to logistic regression and k-nearest neighbors baselines, using individual question features as predictors yielded only a small improvement. This suggests that while overall ability level is the strongest predictor of performance, minor personalization gains are possible by tailoring to exact strengths. Crucially, students did not appear to develop strongly differing ability across topics, undermining the archetype approach. The work offers a new benchmark for the field, demonstrating how explainable models can reach competitive performance on educational data at scale.