The research team, led by Saiful Islam Sagor and Erina Baynojir Joyee of UNC Charlotte’s Department of Mechanical Engineering and Engineering Science, together with colleagues in electrical and computer engineering and an industry partner at Corvid Technologies, set out to answer a question that has become urgent across engineering disciplines. General-purpose models such as GPT-4 and LLaMA have absorbed vast amounts of internet text, yet additive manufacturing knowledge lives in places those models rarely see in depth: peer-reviewed papers, manufacturer documentation, ASTM and ISO technical standards, process guidelines and material datasheets. Earlier studies had already flagged the problem, finding that GPT-4 offered only high-level, superficial responses on manufacturing questions, and that domain-specific systems built around retrieval delivered far more accurate answers.

To run a fair comparison, the researchers first built a curated knowledge corpus of remarkable breadth. It includes more than 250 peer-reviewed research articles, over 30 book chapters, more than 150 additive manufacturing process guidelines, over 50 international technical standards and standard operating procedures, and more than 120 material datasheets. Four researchers manually reviewed and filtered every document, removing duplicates and low-quality material. The collection spans the major families of 3D printing processes, from fused deposition modeling and stereolithography to direct ink writing, directed energy deposition, binder jetting, two-photon polymerization and even 4D printing with shape-memory materials, along with subdomains such as material science, machine settings, design rules and failure analysis.

All three model configurations in the study were built on the same foundation: LLaMA-3-8B, an open-weight, eight-billion-parameter transformer model chosen for its balance of capability, computational accessibility and reproducibility. The baseline configuration used the pretrained model as-is, answering questions with only the general knowledge learned during pretraining. The fine-tuned version was adapted using QLoRA, a parameter-efficient technique that compresses the model to 4-bit precision and inserts small trainable low-rank matrices into the attention layers, allowing domain adaptation on a single GPU-scale budget. Training ran for nine epochs with the AdamW optimizer and early stopping based on validation loss, and the loss curves converged cleanly, indicating the poor later performance was not caused by overfitting.

The retrieval-augmented generation system took a fundamentally different approach. Rather than baking knowledge into the model’s weights, it stores the curated corpus as dense semantic embeddings in a FAISS vector database, built with sentence-embedding models from the Sentence-Transformers framework. At inference time, each user question is embedded and matched against the index to pull out the most relevant document chunks, which are then appended to the query as context. A prompt template explicitly instructs the model to ground its answer solely in the retrieved material, a constraint designed to suppress hallucination. The architecture mirrors the logic of DeepMind’s RETRO model, which matched GPT-3 performance with 25 times fewer parameters by leveraging large-scale retrieval.

Evaluation rested on a benchmark of 200 expert-designed questions spanning fundamental concepts such as typical layer heights in fused deposition modeling, intermediate queries about process and material-property relationships, and advanced problems in troubleshooting, parameter optimization and defect formation. Experts assessed responses in blinded pairwise comparisons, judging accuracy, relevance, harmfulness, contextual correctness, understandability and real-world applicability. The verdict for retrieval was emphatic: retrieval-augmented responses were judged more accurate than the baseline in 75.5 percent of comparisons, more relevant in 90.8 percent, and preferred overall in 85.2 percent. Every one of the retrieval system’s outputs was rated non-harmful, 91.3 percent were clearly understandable, and 74 percent offered guidance usable in real engineering contexts.

The fine-tuned model’s numbers told the opposite story. Trained on raw, unstructured domain text with a causal language modeling objective, the model produced more accurate answers than the baseline in only 5.6 percent of comparisons, and more relevant answers in just 32.5 percent. Only 8.5 percent of its responses were considered suitable for real-world use, and just 31.8 percent were judged understandable. The authors attribute this not to overfitting, which the loss curves rule out, but to a mismatch between the training objective and the task: predicting the next token from raw text teaches a model domain vocabulary, but it never teaches the model how to answer a structured technical question accurately or concisely. Exposure to knowledge, in other words, is not the same as learning to use it.

The implications extend well beyond 3D printing. Fine-tuned models are static; incorporating new standards, new materials or new process data means retraining, and their answers offer no transparent link to sources. Retrieval systems, by contrast, can absorb new documents simply by updating the index, and every answer can be traced back to retrieved reference material, a property that matters enormously in fields where correctness and traceability are non-negotiable. The findings align with a growing body of evidence from medicine, biomedical natural language processing and manufacturing showing that hybrid architectures pairing language models with external knowledge retrieval outperform parameter-based adaptation for knowledge-intensive tasks.

The authors are careful to note the limits of their conclusions. The benchmark contains 200 questions, which cannot capture the full diversity of real-world queries; expert judgment, though structured and blinded, introduces subjectivity; and the fine-tuning experiment used raw text rather than supervised question-answer pairs, which better instruction tuning might have improved. The study also tested only one model family at one size, leaving open whether larger or proprietary models would shift the balance. The researchers propose several next steps, including supervised fine-tuning on high-quality question-answer pairs, hybrid approaches combining retrieval with instruction tuning, comparisons against state-of-the-art proprietary systems, and integration of real-time process monitoring signals such as thermal and optical inspection data to support live manufacturing decisions.

Even with those caveats, the central message lands with unusual force for an AI benchmark study. For engineers hoping to deploy trustworthy AI assistants in specialized technical domains, the cheapest and most effective route may be to stop trying to stuff knowledge into model weights and instead give the model a well-organized library to consult. In additive manufacturing, where a single wrong parameter recommendation can warp a part, ruin a build or compromise a component, the difference between 85 percent and 5.6 percent is not a statistic. It is the difference between a tool an engineer can trust and one they cannot afford to use.

Subject of Research: Domain adaptation of large language models for additive manufacturing question answering

Article Title: Domain adaptation of large language models for additive manufacturing using retrieval-augmented generation and fine-tuning

Article References: Domain adaptation of large language models for additive manufacturing using retrieval-augmented generation and fine-tuning. (n.d.). https://doi.org/10.3389/fmtec.2026.1875357

Image Credits: AI Generated

DOI: 10.3389/fmtec.2026.1875357

Keywords: large language models, retrieval-augmented generation, fine-tuning, additive manufacturing, 3D printing, LLaMA-3-8B, QLoRA, vector database, expert evaluation, artificial intelligence, manufacturing technology, hallucination