Adaptive Question Generation and Dynamic Difficulty Scaling Using Large Language Models for Personalized Assessment

This study proposes and validates an adaptive assessment framework that integrates controlled question generation and dynamic difficulty regulation using Large Language Models (LLMs). The system expands four mathematical domains (Arithmetic, Algebra, Geometry, and Statistics) into an 1,840-item bank produced from 460 curated seed prompts. Iterative LLM refinement increased accepted question rates from 63 percent in the first-generation cycle to 84 percent by the fifth cycle, with rejection rates declining from 37 percent to 16 percent as prompt constraints and screening improved. Human expert evaluation recorded mean scores of 4.3 for clarity, 4.5 for topical relevance, and 4.1 for difficulty appropriateness on a 5-point scale, while originality received 3.8 due to controlled structural similarity. Behavioral testing involving three adaptive rounds produced a mean learner accuracy improvement from 0.64 to 0.79, confirming the effect of ability-matched sequencing. Difficulty tier validation showed mean accuracies of 0.87, 0.71, and 0.52 for Easy, Medium, and Hard questions, respectively, demonstrating differentiated challenge levels. Overall, findings confirm that LLM-driven generation coupled with behavioral difficulty scaling can produce psychometrically reliable, computationally efficient, and pedagogically aligned adaptive assessments suitable for real-time deployment.