Researchers map LLM approaches for automated Arabic text scoring

A new literature review from researchers at an Algerian institution systematically examines how large language models (LLMs) are being applied to automatically score Arabic text in educational settings. The paper, accepted at the 2026 NCMAI conference, categorizes existing work into short answer grading (ASAG) and essay scoring (AES), two distinct tasks with different challenges. The authors introduce a novel taxonomy with five dimensions: application domain (what type of text is scored), feedback generation capability (whether the system provides explanations alongside scores), LLM architecture deployed (from smaller fine-tuned models to large API-based ones), alignment with competency referential frameworks (e.g., CEFR for languages), and prompt engineering strategy (zero-shot, few-shot, chain-of-thought).

After applying this taxonomy to compare studies, the authors find that Arabic ATS remains underexplored compared to English, despite the growing availability of Arabic-specific LLMs and datasets like ARABIC-AQA. Most existing work relies on fine-tuned encoder models rather than generative LLMs, and few systems provide actionable feedback. The review calls for more pedagogically grounded research that incorporates competency frameworks and prompt engineering to improve educational quality across Arabic-speaking communities. This structured analysis helps educators and developers understand where to focus efforts for scalable, consistent evaluation of Arabic learner responses.