Generative AI for Early Grade Story Generation Using a Self-Reflective Approach

Reading comprehension is a crucial skill for developing students' language, insight, and reasoning skills. However, creating high-quality content for large-scale reading assessments is costly and time-consuming. For these assessments to be effective, stories must meet standards such as those defined by the Early Grade Reading Assessment (EGRA), which emphasize structure, readability, clarity, and age-appropriateness. Although many Natural Language Processing (NLP) techniques and Large Language Models (LLMs) have been used for automated story generation, achieving both EGRA compliance and narrative diversity remains challenging. Our paper proposes a solution to increase adherence to the EGRA criteria while ensuring narrative diversity by utilizing GPT-4o to generate children's reading comprehension stories, guided by a database of classic tales. We present a systematic autoregressive framework that incorporates self-reflection to enhance and refine the quality of the stories generated. The stories are evaluated through both automated metrics and human assessment. Our results demonstrate GPT-4o's potential to improve resource efficiency and scalability in literacy assessment practices, providing practical solutions for educators and policymakers in early grade learning.