Children's English Reading Story Generation via Supervised Fine-Tuning of Compact LLMs with Controllable Difficulty and Safety

Problem
This preprint addresses the gap in the capability of Large Language Models (LLMs) to generate age-appropriate children’s stories that align with specific reading levels and safety standards. Existing models, such as GPT-4o and Llama 3.3 70B, often produce narratives that are too complex for the target demographic, limiting their utility in educational contexts. Additionally, the high operational costs associated with large models hinder their adoption in schools and homes. The authors propose a solution that leverages compact LLMs fine-tuned for controllable difficulty and safety, aiming to enhance accessibility and effectiveness in children’s literacy development.

Method
The authors fine-tuned three 8B-parameter LLMs using a dataset derived from an expert-designed children’s reading curriculum and stories generated by larger models (GPT-4o and Llama 3.3 70B). The fine-tuning process involved supervised learning techniques that emphasized controllability over model scale. The training data included various difficulty levels and error patterns, allowing the models to generate stories tailored to specific reading competencies. The evaluation of the generated stories was both quantitative, focusing on difficulty-related metrics, and qualitative, assessing narrative engagement and safety. The training compute details were not disclosed, but the emphasis was on achieving effective performance with a compact architecture.

Results
The fine-tuned 8B LLMs outperformed the zero-shot outputs from GPT-4o and Llama 3.3 70B on difficulty-related metrics, demonstrating a significant improvement in generating age-appropriate content. The authors report that the fine-tuned models produced stories that were more suitable for children’s reading levels, with a marked reduction in complexity. Additionally, the generated stories exhibited almost no discernible safety issues, indicating that the models can be safely deployed in educational settings. Specific effect sizes were not provided, but the qualitative assessments suggest a strong preference for the fine-tuned outputs among educators and parents.

Limitations
The authors acknowledge that their study is limited by the scope of the fine-tuning dataset, which may not encompass the full diversity of children’s reading preferences and styles. Furthermore, the reliance on a single expert-designed curriculum may introduce biases that affect the generalizability of the results. The models’ performance in real-world settings, particularly in varied educational contexts, remains to be validated. Additionally, the computational resources required for fine-tuning, while lower than those for larger models, may still pose a barrier for some users.

Why it matters
This work has significant implications for the development of educational tools that leverage AI for literacy. By demonstrating that compact LLMs can effectively generate engaging and age-appropriate stories, the authors pave the way for broader adoption of AI in educational settings. The focus on controllability and safety addresses critical concerns for educators and parents, potentially leading to enhanced learning outcomes for children. This research could inspire further exploration into fine-tuning methodologies for other educational applications, expanding the role of AI in personalized learning experiences.

Authors: Qian Shen, Fanghua Cao, Min Yao, Shlok Gilda, Bonnie J. Dorr, Walter L. Leite
Source: arXiv:2605.13709
URL: https://arxiv.org/abs/2605.13709v1