Researchers from Johns Hopkins University have identified that 'monofacts'—data points appearing exactly once in a model's training set—significantly enhance the ability of large language models to make scientific discoveries. The study, titled 'Tracing the Scientific Discovery Capacity of LLMs to Their Training Data,' was published on November 3, 2026.
The research team tested this hypothesis in the fields of materials science and regulatory genomics. Models were evaluated on their ability to identify DNA sequences and materials that were held out from training and emerged after the model's training cutoff date. This methodology was designed to ensure that the models were performing genuine discovery rather than simply recalling memorized information.
The findings indicate that as the rate of monofacts increases, the capacity for scientific discovery rises reliably. The researchers noted that this improvement cannot be explained solely by the overall diversity of the training data. The study involved multiple open-weight model families, various sampling strategies, and different reasoning budgets to verify the consistency of the results.
The team observed that the discovery effect can transfer between scientific domains, provided that the training data contains relevant in-domain examples. This research builds upon previous work from June 2026, which explored the emergence of in-context learning capabilities in genomic models trained on non-language sequences like DNA.
Co-authored by researchers from Microsoft and Apple, the study suggests that data curation strategies may need to be adjusted. If monofacts are indeed drivers of discovery, the decision-making process regarding which data to include, remove, or duplicate in training sets becomes a critical factor for future AI development in scientific fields.
