Generating Reading Assessment Passages Using a Large Language Model

The growing demand for high-quality items in computer-based assessments has made the item creation process costly and labor-intensive, relying heavily on human expertise. While automated item generation has been around, large language models can enhance efficiency and quality. In this study, we explored the use of GPT family models to generate reading passages for the Progress in International Reading Literacy Study (PIRLS). Creating passages for 4th graders requires careful attention to complexity, engagement, and relevance. By using well-designed prompts, we generated multiple passages and selected those closely matching original texts based on Lexile scores. All AI-generated passages along with original passages are evaluated by human judges according to their coherence, appropriateness to 4th graders, and readability.