A Johns Hopkins University study shows that facts appearing only once in training data, called “monofacts,” improve the ability of large language models to make new scientific discoveries. Researchers tested the relationship in materials science and regulatory genomics, using materials and DNA sequences that had never been seen during training. As the proportion of monofacts increased, the discovery rate rose reliably. The effect persisted across several open-weight model families, multiple sampling strategies and a range of reasoning budgets, and could not be explained by training-data diversity alone. Transfer between scientific domains worked only when the training data included relevant in-domain examples.
Source: Read the original article
