Can LLMs Control Readability? A Multi-Dimensional Evaluation Framework for CEFR-Controlled Arabic Generation

This study introduces a multi-dimensional evaluation framework for generating Arabic text controlled by the Common European Framework of Reference (CEFR), focusing on the reliability of LLMs in managing readability levels. The framework employs controlled prompting, automatic readability prediction using the Taha-19 model, and syntactic complexity profiling, demonstrating that CEFR-guided prompting significantly enhances alignment with linguistic profiles and readability predictions. These results provide a foundation for incorporating readability-aware text generation in adaptive language learning systems, which is crucial for practitioners developing educational AI applications.