A new study reveals a direct connection between the information composition of training datasets and the ability of large language models (LLMs) to make scientific discoveries. Researchers from Johns Hopkins University and other institutions found that the rate of "monofacts", facts appearing only once in a dataset, within a model's training data is a key determinant of its discovery-making potential. This finding suggests that the unique or rare pieces of information encountered during training are more critical for novel insights than the overall diversity of the data.
The study approached the question of LLM scientific discovery through the lens of "missing probability mass," a theoretical framework that examines how models estimate probabilities for statements not explicitly seen in training. The researchers developed a theory positing that monofacts set the capacity for discovery. To test this, they conducted experiments in materials science and regulatory genomics. In these experiments, they directly altered the proportion of monofacts in the training sets used for two different open-weight LLM families.
The results demonstrated a consistent pattern: datasets with a higher rate of monofacts yielded more discoveries across various sampling strategies and computational budgets. These findings were not explained by training set diversity alone. Further experiments indicated that this effect could transfer across different scientific domains when paired with domain-specific data, suggesting a general principle at play.
This research builds upon previous work that explored the relationship between training data characteristics and LLM behavior, particularly concerning "hallucinations", the generation of plausible but false information. Studies have shown that hallucination rates are also linked to the monofact rate, suggesting that the long tail of the training distribution, where rare facts reside, influences both novel discovery and factual inaccuracies. The current work extends this by focusing specifically on the positive aspect of discovery rather than just error reduction.
The experiments involved tasks such as training LLMs to generate chemical formulas that meet specific material properties and proposing DNA sequence modifications for desired regulatory activity. Discovery in these contexts was measured by the models' ability to accurately predict outputs for data that occurred after their knowledge cutoff, effectively requiring them to generalize beyond their training experience.
The implications of this study could influence how large datasets for training AI models are curated. Historically, practices like deduplication have aimed to clean datasets, but this research suggests that the sheer uniqueness of information, represented by monofacts, might be more valuable for fostering discovery capabilities than previously understood. This challenges the assumption that simply increasing data volume or diversity is the sole path to enhancing LLM scientific reasoning. The work provides a quantitative link between data composition and an LLM's capacity for scientific exploration, offering a new perspective for developing AI systems that can contribute meaningfully to scientific advancement.