Enhancing Automated Feedback in Intelligent Tutoring Systems via Large Language Models and Data Distillation

Large Language Models (LLMs) have shown promise in generating real-time, adaptive feedback in Intelligent Tutoring Systems (ITS), but their effectiveness is limited by hallucinations, misconception reinforcement, fairness disparities, and high computational cost. This study introduces a structured data distillation pipeline to refine LLM-generated feedback, improving accuracy, fairness, and efficiency. Using a synthetic student simulation (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{N}=100,000$</tex>), we evaluate the impact of different feedback mechanisms on learning gains, retention, and misconception correction, applying probabilistic cognitive modeling, Bayesian adaptive task allocation, and causal inference. Results show that LLM+DD (LLM with Data Distillation) significantly reduces misconception propagation (72 %), mitigates fairness disparities (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$4 4 \%$</tex>), and lowers inference cost (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3 7 \%$</tex>) compared to standard LLM-based feedback, approaching the performance of human expert feedback. A multi-objective optimization analysis further quantifies the trade-offs between pedagogical effectiveness and computational feasibility, demonstrating that LLM+DD enables scalable ITS deployment without sacrificing response quality. These findings establish a structured framework for refining AI-powered tutoring feedback, ensuring LLM-based ITS models align with human pedagogical standards while maintaining efficiency for widespread adoption.