Abstract:

Teachers working with students with special educational needs (SEN) in lower and upper secondary education and vocational pathways face a persistent tension: they are expected to provide accessible learning materials tailored to diverse profiles, while operating under heavy workload, limited time and scarce specialist support. Preparing easy-to-read texts, reformulating complex task descriptions, breaking assignments into short, sequenced steps, and adding explicit metacognitive prompts are all recognised good practices for inclusion, but they are also labour-intensive and often left to individual goodwill rather than supported by systemic tools and workflows.

In parallel, the rapid diffusion of generative artificial intelligence (AI), especially large language models (LLMs), has created both expectations and concerns in education. While mainstream LLM-based tools can assist with generic text rewriting or summarisation, they are rarely optimised for SEN contexts, are typically large and computationally expensive, and tend to operate as opaque “black boxes” with limited pedagogical control. There is a clear gap between generic AI writing tools and the specific, fine-grained accessibility requirements that inclusive education demands.

This paper presents the design of "Micro-LLM for SEN", a research and development project that aims to prototype a lightweight, teacher-centred AI assistant to support the authoring of accessible learning materials for SEN students.

The project is grounded on three design principles:
(1)-computational efficiency, by relying on small or medium-sized language models and parameter-efficient fine-tuning techniques so that the assistant can run on modest hardware;
(2)-pedagogical transparency and control, by encoding accessibility guidelines (e.g., easy-to-read criteria, task scaffolding patterns, tone and vocabulary constraints) into prompt templates, rubrics and guardrails that remain visible and editable for teachers; and
(3)-openness and adaptability, enabling schools to customise the assistant to their subjects, curricula and local support practices through modular datasets and adapters.

The paper adopts a design-based research perspective and describes the methodological framework of the project rather than final impact results. It details the planned phases of corpus construction and annotation, co-design workshops with teachers and counsellors, and a quasi-experimental study that will compare AI-assisted and traditional authoring workflows. Key outcome measures include time required to produce adapted materials, perceived workload, perceived usefulness and control, and expert ratings of clarity, accessibility and alignment with SEN-oriented criteria.

Expected outputs include:
(a) a reusable workflow for AI-assisted accessible authoring;
(b) a set of open rubrics, prompt templates and examples that operationalise accessibility guidelines for SEN in a form usable by teachers and AI systems; and
(c) empirical evidence on the opportunities and limitations of lightweight generative AI as a “copilot” for inclusive material design.

By sharing the project design and methodological choices at an early stage, the paper seeks to inform and invite dialogue within the educational technology community about realistic, ethically grounded ways of integrating generative AI into teachers’ everyday work for inclusion.

Keywords:

Generative artificial intelligence, Lightweight language models, Special educational needs (SEN), Accessible learning materials, AI-assisted authoring, Inclusive education, Teacher workload, Co-design with teachers.