Children's English Reading Story Generation via SFT of Compact LLMs

Key Finding: Small Fine-Tuned Models Beat Large Zero-Shot Models

Using an expert-designed children's reading curriculum and stories generated by GPT-4o and Llama 3.3 70B as training data, the authors fine-tuned three different 8B-parameter LLMs. The fine-tuned 8B models outperformed zero-shot GPT-4o and Llama 3.3 70B on difficulty-related metrics while showing negligible safety issues.

The Controllability-over-Scale Paradigm

The core philosophy is controllability over scale: a compact, affordable model can be fine-tuned to target specific reading levels and error patterns, something large general-purpose models cannot do out of the box. This has major implications for Equity In AI Education — schools could run these models locally at low cost rather than depending on expensive API-based LLMs.

Educational Deployment

The generated stories were designed for use by teachers, parents, and children in classrooms and at home. The controllable difficulty enables Personalized Learning at scale — matching reading materials to individual student proficiency levels. The safety guarantees address concerns in AI Tutor Safety Harms.

Methodological Contribution

Fine-tuning designs were systematically compared, with the curriculum-derived training data providing curriculum alignment that general-purpose models lack. The quantitative and qualitative evaluation framework provides a template for assessing AI-generated educational content.

Connected Concepts

Connected Articles

Citation

Shen, Q., Cao, F., Yao, M., Gilda, S., Dorr, B. J., & Leite, W. L. (2026). Children's English reading story generation via supervised fine-tuning of compact LLMs with controllable difficulty and safety. Proceedings of the 21st Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2026). arXiv:2605.13709.