| C. Papakostas, C. Troussas, A. Krouska, Ph. Mylonas, C. Sgouropoulou |
| Large Language Models for Adaptive and Intelligent Tutoring Systems: A Systematic Mapping Study |
| 21st International Workshop on Smart Media Adaptation, AI & Personalization (SMAP 2026), November 19-20, 2026, Mystras, Greece |
|
ABSTRACT
|
| Large Language Models (LLMs) are being integrated into Intelligent Tutoring Systems (ITS) faster than any consolidated account of the field has been able to follow. This paper reports a systematic mapping study of 86 primary studies published between 2023 and 2026, identified through a search of Scopus that returned an initial pool of 131 records and screened against explicit inclusion and exclusion criteria. Following established mapping-study practice, the studies are classified from their titles, abstracts and keywords using an inductively derived scheme, and the analysis reports category frequencies rather than a pooled synthesis of outcomes. Four research questions are addressed, concerning architectural approaches, personalisation strategies, the extent of reported evaluation, and open challenges. The map identifies five dominant technical approaches: retrieval-augmented generation (11.6% of studies), fine-tuning (11.6%), knowledge tracing (9.3%), knowledge graph integration (8.1%) and multi-agent frameworks (7.0%). Adaptive behaviour is claimed by 94.2% of the mapped systems and personalisation by 68.6%, yet fewer than one study in ten couples the model to a formal learner model and only 10.5% report an evaluation with human participants. We characterise the challenges named across the corpus, namely hallucination, pedagogical control, evaluation fragmentation, privacy and equity, and derive from the map a prioritised six-point research agenda. The major conclusion of this research is that the binding constraint on the field is no longer technical capability but pedagogical grounding and empirical validation: adaptivity is claimed almost universally, whereas formal learner modelling and evaluation with learners remain rare, and closing that gap, rather than adding architectural variety, is what will determine whether LLM-based tutoring delivers measurable learning gains. To the best of our knowledge, this is the first study to map the LLM-based ITS landscape jointly across architectural, pedagogical and evaluation dimensions.
|
| 19 November , 2026 |
| C. Papakostas, C. Troussas, A. Krouska, Ph. Mylonas, C. Sgouropoulou, "Large Language Models for Adaptive and Intelligent Tutoring Systems: A Systematic Mapping Study", 21st International Workshop on Smart Media Adaptation, AI & Personalization (SMAP 2026), November 19-20, 2026, Mystras, Greece |
[
BibTex] [
Print] [
 Back] |
