Pedagogical Alignment, Adaptivity, and Analytics in Intelligent Tutoring Systems: A Systematic Review

Abstract

Intelligent tutoring systems have changed quickly since large language models became widely available in 2023, raising a practical question for e-learning research: when a tutor is built on a general-purpose language model rather than on hand-encoded rules, what happens to its pedagogical grounding, learner adaptivity, and reproducibility infrastructure? Earlier reviews have examined these matters separately, but none has considered how they come together in the systems now being built. This systematic review addresses that gap through three analytical axes: pedagogical alignment with an instructional theory, adaptivity and learner modelling, and analytics standards supporting comparison and reproducibility. The review followed the SPAR-4-SLR protocol and used PRISMA 2020 as a reporting framework where the recoverable record allowed. Candidate records were screened, verified against their source documents, and coded conservatively against pre-specified criteria. Twelve primary studies were retained after source verification. Pedagogical alignment and adaptivity were each addressed by eleven of the twelve studies (92 percent), although implementation varied widely, from explicit frameworks such as Cognitive Apprenticeship, Productive Failure, and Socratic questioning to looser prompt-based behaviour. Analytics standards and reproducibility were the weakest axis, addressed by three studies (25 percent). No study combined all three axes with explicit uncertainty quantification or calibration of tutoring decisions. Controlled learning-outcome evidence was also scarce: only one study reported a controlled comparison, and its findings were preliminary. For e-learning practice, the review gives teachers, platform designers, and institutional decision-makers a realistic basis for expectation: current language-model tutoring systems are promising and often pedagogically plausible, but their learning benefits remain under-demonstrated and should be evaluated locally before large-scale adoption. For e-learning knowledge, the review contributes a source-verified synthesis of recent intelligent tutoring research and identifies a clear methodological gap: the field is progressing faster in system design than in shared evaluation, interoperable analytics, and calibrated decision-making. The paper concludes that future work should combine instructional theory, inspectable learner models, standardised logging, open benchmarks, and uncertainty-aware evaluation within the same tutoring systems.