RAG beats fine-tuning: LLaMA-3-8B excels in polymer 3D printing Q&A

A new arXiv paper (2605.12516) from researchers at [University not specified] tackles a persistent problem: general-purpose LLMs like LLaMA-3-8B struggle with specialized engineering domains where answers require precise technical knowledge from dispersed sources. The team focused on polymer-composite additive manufacturing (3D printing), where information lives in academic papers, manufacturer specs, standards, and procedural guides. They tested three configurations: the pretrained baseline, a RAG system retrieving relevant document chunks from a vector database, and a model fine-tuned on raw domain text. Evaluation used 200 expert-written questions judged by mechanical engineering PhDs for accuracy, relevance, and overall preference.

The results are striking. RAG dramatically outperformed both the baseline and fine-tuned models: 75.5% of RAG answers were judged more accurate, 85.2% were preferred overall, and 90.8% were rated more relevant than baseline responses. In contrast, fine-tuning on unstructured AM text backfired — only 5.6% of fine-tuned answers were more accurate, and just 32.5% were more relevant. The study underscores that for technical fields with specialized, scattered knowledge, retrieval-augmented generation provides a far more practical and reliable path than naive fine-tuning, which can degrade model performance on raw, noisy data.