FROM SYMBOLIC RULES TO LANGUAGE MODELS: A SYSTEMATIC LITERATURE REVIEW OF THE TECHNIQUE-LEVEL EVOLUTION OF INTELLIGENT TUTORING SYSTEMS

ABSTRACT

Intelligent tutoring systems (ITS) have evolved through four technical eras: symbolic, rule-based expert systems; Bayesian and statistical student modelling; deep learning knowledge tracing; and, most recently, LLM-based conversational tutors. Existing reviews examine these eras separately, or discuss the history at a pedagogical rather than technique level, without one consistent comparative framework. Following an adapted PRISMA 2020 protocol, foundational and representative papers were identified across Google Scholar, Semantic Scholar, arXiv, and the ACM/IEEE digital libraries. Twenty-eight verified sources were synthesised into an explicit four-era taxonomy, compared along seven dimensions: knowledge representation, student modelling mechanism, adaptivity logic, feedback generation, data/engineering burden, interpretability, and effectiveness evidence. Results show that ITS development has not been linear: each era gains capability—chiefly scalability and natural-language competence—while losing interpretability. The review concludes that future work should combine LLM-based generation with the validated, interpretable student modelling of Bayesian systems, and proposes a research agenda for neuro-symbolic hybrid tutors.

Keywords: Intelligent tutoring systems; knowledge tracing; Bayesian Knowledge Tracing; deep knowledge tracing; large language models; conversational tutors; systematic literature review; taxonomy