Intelligent Tutoring Systems (ITS) aim to provide adaptive learning experiences through personalized scaffolding during the inner loop. However, automating this process remains challenging due to limitations in dynamic adaptability and over-reliance on direct solutions. This study addresses these gaps by leveraging Large Language Models (LLMs) to generate and integrate multi-step scaffoldings into the ITS inner loop. Grounded in relevant educational theories, we propose a framework for generating structured scaffoldings using LLMs, which could guide learners through problem-solving stages (representation, planning, execution, evaluation) while fostering meta-cognitive skills. A dual-module architecture—comprising a Scaffolding Generation Module (integrating LLMs for automated scaffolding creation) and a Scaffolding Integration Module (managing adaptive learner-system interactions)—was developed and evaluated. Empirical experiments with 28 middle school students demonstrated significant learning gains (pre-test: 0.23 vs. post-test: 0.68, p < 0.000), while fsQCA revealed distinct paths to success, highlighting the critical roles of execution scaffolding for high-prior-knowledge learners and holistic scaffolding for low-prior-knowledge learners. Our findings advance the integration of LLMs in ITS by balancing accuracy, adaptability, and pedagogical efficacy, offering a scalable approach to fostering deep thinking in real-time learning environments.