An AI assisted four layer teaching model for computational thinking cultivation evaluated in a 14 week classroom pilot

Abstract

Computational thinking has become a foundational competency for 21st-century learners, yet effective and scalable approaches to its cultivation remain insufficient. This study addresses three persistent gaps in the literature: the lack of systematically integrated AI-assisted pedagogical architectures, the inadequate operationalization of personalized learning paths for computational thinking, and the reliance on single-dimensional evaluation metrics. We propose a four-layer AI-assisted teaching model that, building on the conventional resource–adaptation–interaction–evaluation lineage of intelligent tutoring systems, attempts to integrate intelligent resource management, adaptive learning path generation based on deep knowledge tracing, multimodal teaching interaction support, and a fuzzy comprehensive evaluation framework spanning five core computational thinking dimensions across three competency tiers. To probe the feasibility of this integrated package, we ran a 14-week quasi-experimental study involving two intact classes ( \(\:n=120\) ) from a single university. Each class received one condition in its entirety, so class membership and treatment condition cannot be separated, and we read the comparison as a two-class pilot rather than as an estimate of a generalizable treatment effect. In this single-site sample, the experimental group scored higher on the composite computational thinking measure than the control group ( \(\:t\left(118\right)=5.36\) , \(\:p<0.001\) , Cohen’s \(\:d=0.97\) ), with the largest effect sizes observed in abstraction ( \(\:d=1.03\) ) and evaluation-with-generalization ( \(\:d=0.98\) ). An ANCOVA adjusting for pre-test scores left the group difference in place, although with a single class per condition no covariate can absorb what operates at the level of the class. The Week 5 and Week 9 assessments used abbreviated parallel forms that were never formally equated to the full instrument, and they are reported as descriptive trajectory indicators. Regression analysis indicated that, among the behavioural correlates collected, task completion volume and debugging efficiency were the strongest predictors of learning gains. Subgroup analyses suggested that medium-experience learners appeared to benefit most from the adaptive sequencing, while low-experience learners gained primarily on foundational dimensions. We read these findings as preliminary evidence that, under our specific implementation and student population, a coherently integrated AI-assisted package can be associated with larger composite gains than conventional instruction; whether the adaptive engine itself, rather than richer resources or longer online engagement, is the operative ingredient remains an open question that an active-control, dismantling study would be needed to settle.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗