Towards automated accessibility adaptation of learning materials with compact large language models

This paper investigates the use of locally deployed Compact Large Language Models (CLLMs) to support inclusive education by automatically adapting learning materials for students with diverse accessibility needs. We propose a structured framework for multi-profile accessibility adaptation based on explicit, constraint-driven transformations, modeling learner profiles (e.g., dyslexia, ADHD, visual impairments) as sets of formalized requirements integrated into a modular pipeline. To support practical use, we also developed a web-based interface that allows users to input educational content, select a learner profile and a language model, and generate adapted versions of the text. We evaluate the framework using three locally executed models: Llama 3 8B, Mistral 7B, and Phi-3 Mini. Results show that larger models better comply with complex constraints, though all models exhibit limitations, including partial adherence to constraints and formatting issues. These findings highlight both the potential and the current limitations of LLM-based approaches for scalable accessibility support in education.