Large Language Models for Reading Passage Generation: A Closed-Loop Approach to Difficulty Adaptation

Abstract

This thesis investigates how large language models can be used to generate reading materials that are usable, controllable, and analysable for graded English reading practice. Traditional approaches to reading material preparation are costly and slow to update, making them illsuited to personalised learning and dynamic difficulty adaptation. Although large language models demonstrate strong capabilities in text generation and rewriting, direct prompting often fails to satisfy target length, target difficulty, and content fidelity constraints in a stable manner. To address this problem, this thesis designs and implements a closed-loop system for reading passage generation and difficulty adaptation, integrating passage generation, readability evaluation, iterative rewriting, result archival, and statistical analysis into a unified workflow. Methodologically, the thesis operationalises reading difficulty as a calculable metric target and models closed-loop adaptation as an iterative Generate-Evaluate-Adjust-Archive-Analyse process. The system first generates an English reading passage according to a specified topic, target level, and target word count. It then evaluates target difficulty using a composite difficulty signal, referred to in this thesis as the grade-family mean, which consists of Flesch-Kincaid Grade, Automated Readability Index, Coleman-Liau Index, and Gunning Fog, while retaining Flesch Reading Ease as a monitoring signal. During the difficulty adaptation stage, the system employs a constrained replacement-only rewriting strategy and constructs a Content Fidelity score from the recall of entities, numbers, and keywords in order to balance difficulty alignment against semantic preservation. In addition, the system supports local data archival, batch experiment logging, and multidimensional statistical analysis, thereby providing a basis for later reproduction and academic evaluation. Two categories of experiments were conducted around the system: direct passage generation and closed-loop difficulty adaptation. The results show that direct prompting can produce broadly usable reading passages, but its difficulty control remains unstable; in the final reporting subset, the maximum hit rate of the direct-generation task was 31.6%. By contrast, closed-loop difficulty adaptation substantially improved target alignment, achieving hit rates from 92.5% to 98.8% within the final experimental scope while maintaining fidelity at approximately 0.92 to 0.93. These findings indicate that, in graded reading scenarios, difficulty control is more effective when reformulated as a feedback-driven iterative optimisation problem rather than a oneshot prompting problem. The principal contributions of this thesis are fourfold: it proposes a closed-loop difficulty adaptation framework for educational text generation; develops an adaptation mechanism that jointly considers difficulty alignment and content fidelity; implements a dynamic reading-material system that is operational, archivable, and analysable; and empirically validates the effectiveness of the proposed approach in the context of reading passage generation.