A Multi-Dimensional Framework for Adaptive Scaffolding in Large Language Model

Abstract

Adaptivity in systems based on large language Models (LLMs) is a relatively new direction in the research of intelligent tutoring systems (ITSs) and user modeling. This paper gives a detailed description of how adaptivity is realized in ITSs that integrate LLMs or LLMs themselves as pedagogical tools. The review aims to answer the question: What are the constitutive dimensions of adaptivity in dialogue-based scaffolding with Large Language Models, and how can these dimensions be formalized into a multi-dimensional framework to guide design and evaluation? We began by defining the concept of adaptivity based on comparison between the implementation of the rigid, rule-based architectures of traditional ITS and probabilistic, generative approaches capable of dynamic adaptation. The field of AIED currently lacks a theoretical framework for conceptualizing adaptivity in LLM-based systems. We address this gap by synthesizing findings from a systematic literature review and coupling them with inductive thematic analysis to develop a comprehensive multi-dimensional framework of adaptivity. We analyze and identify four constitutive dimensions through which LLM-based systems modify their behavior in response to learner needs: Cognitive Adaptivity (modulating informational content to match knowledge states), Strategic Adaptivity (adjusting pedagogical structure and timing of support), Affective Adaptivity (regulating emotional engagement through tone and persona), and Metacognitive Adaptivity (supporting selfregulation and reflective learning processes). All dimensions proposed in the analysis phase are grounded in learning sciences theory, particularly Vygotsky’s Zone of Proximal Development and Wood, Bruner, and Ross’s concept of scaffolding. We also discuss the implications of this shift, including the hallucination trade-off and ethical considerations surrounding overscaffolding. Providing a framework geared toward designers and researchers for designing pedagogically effective LLMbased tutoring systems that move beyond superficial conversational fluency toward strong educational impact.