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

Their system, called React-to-Me, is a conversational assistant built on top of the Reactome Pathway Knowledgebase, one of the most rigorously maintained curated resources in molecular biology. Reactome, stewarded by a team of expert biologists, documents human biological pathways, molecular interactions, and disease mechanisms with a level of care that ordinary web search cannot match. Yet the knowledgebase has a usability problem: its data model is intricate, and its search interface rewards people who already know the vocabulary and structure of the underlying database. For a graduate student, a clinician, or a researcher crossing into a new field, that learning curve can be discouraging. React-to-Me was designed to let such users ask questions in plain natural language while still receiving answers that are traceable, verifiable, and anchored in curated content.

The engineering heart of the system is a hybrid retrieval-augmented generation pipeline, a technique that constrains a language model to answer only from documents fetched at query time rather than from its internal memory. In React-to-Me, an incoming question triggers two complementary searches over the Reactome corpus. The first is a semantic vector search, which uses neural embeddings to match the meaning of the question against the meaning of database entries, capturing paraphrases and conceptual overlap that keyword matching would miss. The second is a lexical keyword search, of the BM25 variety familiar from classic information retrieval, which excels at pinning down precise identifiers, gene names, and chemical terms. The two ranked lists are merged with reciprocal rank fusion, a method that rewards documents appearing near the top of both lists, and the fused evidence is handed to a language model that is instructed to generate an answer grounded strictly in that retrieved material.

Grounding alone is not enough; the system also insists on provenance. Every response is directly linked to the corresponding Reactome entries from which it was assembled, so users can click through and inspect the curated pathways behind the answer. And when the retrieved evidence is thin or absent, React-to-Me is designed to defer rather than improvise: it declines to speculate and instead points users toward trusted external biomedical sources. This refusal behavior, the authors argue, is a feature rather than a failure, because an honest admission of limited coverage is far more valuable to a scientist than a fabricated pathway. The team tested how well that refusal logic worked in the wild by manually classifying logged instances where the system declined to answer, finding that out-of-scope biology questions were the largest category, accounting for 31.9 percent of refusals.

Before any human saw the system, the researchers put the retrieval engine itself through computational benchmarking using the Ragas evaluation framework, scoring configurations on context utilization, relevance, and faithfulness across several query types: simple factual lookup, mechanistic reasoning, multi-context synthesis, and conditional logic. The results showed a clear pattern. Semantic retrieval on its own preserved relevance but faltered on faithfulness, letting unsupported statements slip through, while pure keyword retrieval weakened on complex reasoning tasks that required synthesizing multiple sources. The combined hybrid approach, ranked by reciprocal rank fusion, delivered the highest and most stable scores across every metric and every query type, confirming that the two retrieval strategies compensate for each other’s blind spots.

Real-world deployment brought a second wave of evidence. Analyzing user queries logged during October 2025 and March 2026, the team found that factual lookup and mechanistic or relational biology questions dominated, together accounting for 56.1 percent of classified logged-in queries. Notably, the mix shifted over time: factual lookups rose from 24.7 percent to 33.7 percent of classified queries, while mechanistic and relational questions declined from 32.7 percent to 21.2 percent, a pattern suggesting that as users became comfortable with the tool, they leaned on it increasingly as a rapid reference rather than a reasoning partner. The study received ethics approval from the University of Toronto Research Ethics Board under protocol number 47192, and all survey participants provided informed consent.

The most demanding test was a blinded head-to-head comparison against a general-purpose model, GPT-4o-mini, judged by ten external molecular biology experts who did not know which system produced which answer. The verdict favored grounding decisively. Grounded React-to-Me responses were more likely to receive higher quality ratings than their ungrounded counterparts, with an overall odds ratio of 2.01. The gains were strongest where scientific rigor matters most: factual accuracy improved with an odds ratio of 2.97, biological specificity with an odds ratio of 2.88, and mechanistic depth with an odds ratio of 1.83. Individual evaluator preferences, recorded in the supplementary data, showed consistent majority favoring of the grounded system across question-level comparisons, an exploratory mixed-effects ordinal regression accounting for variation among evaluators and questions.

User sentiment reinforced the expert verdict. In a survey of 25 consenting users drawn from the publicly deployed system, 92 percent expressed strong satisfaction with ease of use, 88 percent with citation reliability, and 85 percent with factual accuracy. Strikingly, perceived accuracy showed the strongest association with overall confidence in the system, with a correlation of r = 0.81, meaning that users who trusted the answers did so primarily because the facts checked out, not because the interface was pretty. Respondents spanned organization types and research fields, with human biology research most represented, followed by computational biology, cell biology, and genetics, and satisfaction held up across education levels and levels of prior Reactome experience.

Beyond the headline numbers, the study offers a practical template for institutions wrestling with how to deploy AI responsibly. The team published monthly operating costs, the full survey instrument, benchmark questions, and analysis code as supplementary materials, an unusual degree of transparency for a production AI deployment. The willingness to defer to trusted sources when coverage runs out, the insistence on clickable citations for every claim, and the hybrid retrieval architecture that balances conceptual breadth with lexical precision together sketch what credible scientific AI might look like. As funders and publishers grapple with hallucination risks in research tooling, React-to-Me demonstrates that domain-specific grounding is not a theoretical nicety but a measurable design choice, one that nearly triples the odds that an expert will judge an answer factually accurate. The broader lesson may extend well beyond pathways: for specialized knowledge, the future of conversational AI belongs to systems that know the boundaries of what they can honestly say.

Subject of Research: A grounded conversational AI interface for querying the Reactome Pathway Knowledgebase

Article Title: React-to-Me: real-world experience of a grounded conversational interface to the Reactome Pathway Knowledgebase

Article References: Mohammadi, H., Almodaresi, F., Hogue, G. F. J., Wright, A., Orlic-Milacic, M., Li, N. T., Mawani, A., & Stein, L. (2026). React-to-Me: real-world experience of a grounded conversational interface to the Reactome Pathway Knowledgebase. BMC Biology. https://doi.org/10.1186/s12915-026-02740-2

Image Credits: AI Generated

DOI: 10.1186/s12915-026-02740-2

Keywords: Reactome Pathway Knowledgebase, conversational AI, retrieval-augmented generation, large language models, knowledge grounding, biological pathways, biomedical question answering, hybrid retrieval, language model evaluation, computational biology, expert evaluation, responsible AI