Pedagogical Efficiency in AI Training: From Reading Recovery to Retrieval Architectures

Recent empirical results from small language models demonstrate that carefully curated,
textbook-quality training data can produce performance comparable to models trained on
orders of magnitude more unfiltered data. This paper proposes that this efficiency gain is not
accidental but reflects well-established principles from educational pedagogy—specifically
Marie Clay’s Reading Recovery methodology. We argue that pedagogical structure, not raw
data scale, is the primary driver of efficient learning in both human literacy acquisition and
artificial language models. Building on this insight, we propose a shift from
generative-dominant architectures toward retrieval-centric systems using Canonical
Knowledge Objects (CKOs), enabling transparent, low-resource, and trustworthy AI
deployment. This reframes AI efficiency as a curriculum design problem rather than a
hardware or scale problem