Comparative analysis of language models in automated educational diagnostic tasks

This study presents a comparative analysis of language models for automated educational diagnosis, evaluating two specialized models (a feature-based ML model and a fine-tuned BERT) against three state-of-the-art Large Language Models (GPT-4, Claude 3, Gemini Pro). Using a mixed-methods approach on a corpus of 450 annotated student texts, the research assessed diagnostic accuracy, feedback quality, robustness, and hallucination rates. Results indicate that advanced LLMs, particularly GPT-4 and Claude 3 Opus, achieve superior diagnostic performance (F1-Score ≥ 0.83) and generate pedagogically higher-quality feedback, rated as more useful and actionable by expert judges. However, these models exhibit a critical weakness: a significant propensity for factual hallucinations (5.2% - 12.3%). In contrast, specialized models, while less sophisticated, offer near-perfect reliability and greater cost-efficiency. The study concludes that LLMs represent a qualitative leap in automated assessment but are not yet reliable for autonomous application. The most viable implementation model is AI-augmented education, where LLMs act as diagnostic assistants to teaching professionals, who provide essential oversight, validation, and personalization. This human-in-the-loop paradigm balances technological potential with pedagogical integrity and ethical responsibility.