The Generalization Gap in the Classroom: An Optimization Lens on Spacing, Variation, and Transfer

Abstract

Educators already picture learning as a climb from recall toward creation. Modern machine learning, it turns out, climbs something that loosely rhymes with that ladder, and the resemblance — imperfect as it is — is enough to ask what the two worlds might teach each other. This is a conceptual synthesis, not an empirical study, and it treats the resemblance as a bounded analogy rather than a claim that human and artificial learning share a mechanism. The genuine parallel is overfitting — performing perfectly on practised material yet failing on novel problems, which is teaching to the test in another vocabulary. We propose overfitting as a bounded organizing idea for a familiar repertoire — spacing, varied practice, desirable difficulties, and transfer testing — around a shared concern: preventing strong performance on familiar items from being mistaken for generalizable learning. So framed, the optimization view introduces no new practices; it offers a complementary account of why the established ones are robust, adding to existing theories of retrieval and transfer an emphasis on how assessment conditions, not learner disposition, come to reward memorization. The framework makes a testable prediction: assessments weighted toward systematically varied, unpractised problems will discriminate genuine competence more sharply than assessments matched to practice, separating students who look equivalent under practice-matched scoring. The prediction applies most clearly to near transfer, where the surface differs but the structure is shared. We are equally clear about the frontier the lens cannot cross: motivation and meaning, which the learner owns in a way no optimized system does.