Optimizing Open Large Language Models for Equity and Accessibility in Higher Education Through Pruning and Retrieval-Augmented Generation

Abstract

This study examines whether open large language models (OLLMs) can be optimized for equitable and efficient use in higher education under resource-constrained conditions. A quantitative experimental design was implemented to evaluate five OLLMs —Falcon, Bloom, GPT-NeoX, T5, and Flan-T5—under four conditions: baseline unoptimized inference (C0), pruning only (C1), retrieval-augmented generation (RAG) only (C2), and pruning combined with RAG (C3). Each condition was tested using 50 educational queries per model across four repetitions. To justify the pruning configuration, an ablation study compared sparsity levels of 10%, 20%, 30%, and 40%, identifying 20% as the best trade-off between efficiency and response quality. The results show that pruning reduced response time by 10.7%, lowered RAM/VRAM usage by 18.9%, and increased throughput by 33.2%. Retrieval augmentation improved educational response accuracy by 1.3 percentage points, with the strongest gains observed for factual queries. Although pruning-only achieved the best efficiency results and RAG-only produced the highest accuracy, the combined condition provided the most balanced profile for realistic educational deployment. These findings suggest that moderate pruning and retrieval augmentation can jointly support lighter, faster, and more contextually grounded language-model deployment in higher education, particularly in institutions with limited computational infrastructure. The study contributes empirical evidence of technical and pedagogical feasibility under simulated deployment conditions.