Tailored Tales: Enhancing Children’s Reading Comprehension with Preference-Tuned Automatic Story Generation

Reading comprehension is crucial for children’s academic success, but creating effective and engaging reading materials on a large scale is challenging. Traditional story creation methods are labor-intensive and often lack flexibility. Large Language Models (LLMs) have emerged as promising tools for automating content generation, but fine-tuning these models to produce high-quality children’s stories can be resource-intensive. Without fine-tuning, LLM-generated stories may not align with the needs of young readers, leading to issues with language, consistency, and educational value. This study explores the use of preference tuning as a cost-effective method for training LLMs to generate diverse, engaging, and age-appropriate stories that meet Early Grade Reading Assessment (EGRA) criteria. Building on previous work, which drew inspiration from classical tales like Arabian Nights and Brothers Grimm, the study uses a dataset of annotated short stories to refine the LLM’s output. The generated stories are evaluated for readability, coherence, and engagement through quantitative measures. Preliminary results show that preference-tuning significantly improves story quality and leads to better comprehension outcomes compared to baseline models.