Design, Integration, and Evaluation of LLM-Enhancement Techniques for LMSs: RAG, LoRA Fine-Tuning, and Structured Generation

Large language models (LLMs) offer significant potential for enhancing learning management systems (LMSs), yet challenges such as outdated information, hallucinations, and non-determinism limit their reliability in learning contexts. This paper develops and evaluates an end-to-end system that addresses these issues through three techniques: retrieval-augmented generation (RAG), low-rank adaptation (LoRA) fine-tuning, and structured (guided) generation. The system integrates directly with LMSs via plugins or a middleware-based architecture, extracts course materials into reusable formats such as Markdown, to be used by different output enhancement techniques. The techniques are evaluated on both public and synthetic learning materials to avoid pretraining bias. Results show that RAG provides the most substantial improvements in performance, particularly when paired with context-aware retrieval. LoRA fine-tuning also improves results, especially with high-quality synthetic data, but its benefits are more limited compared to RAG. Structured generation eliminates schema errors and ensures machine-readable generation of interactive or LMS-native content, such as personalized quizzes. Together, these methods demonstrate how LLM-based systems can generate course-aligned content that is accurate and easily usable with existing LMS platforms. The findings recommend RAG as the primary strategy when a curated corpus is available and LoRA as a viable alternative when RAG is impractical. In addition, structured generation is found as essential for reliability when generating format specific content. This work contributes both a technical framework and applied evidence for integrating advanced LLM systems into LMS environments to support scalable and high-quality AI-enhanced digital learning.