Increasing LLM monofacts raises accuracy… hallucination increases in materials science same-distribution evaluation

Johns Hopkins-led researchers reported that fine-tuning open-weight LLMs gpt-oss-20b and gemma-3-12b-it on materials-science and regulatory-genomics data with a higher share of "monofacts"—facts seen only once in training—improved recovery of unseen correct answers on both tasks. However, on some materials-science evaluations the same approach also increased hallucinations, or incorrect answers. The work, released as a public preprint rather than a peer-reviewed journal article, defines monofacts as training examples observed exactly once and tracks how their frequency in fine-tuning data affects answer accuracy. The findings matter because they suggest that upweighting rare, single-exposure facts can boost generalization but at a cost in reliability for scientific question-answering. What to watch next is whether the accuracy-hallucination trade-off holds across other models, domains and evaluation sets, and whether data-mixing ratios can be tuned to capture the gains without the added errors.