Large language models (LLMs) can propose scientific hypotheses, but it is unclear how training data allows them to make correct discoveries instead of generating plausible falsehoods. A key factor is how often each fact appears in training: facts seen only once, called monofacts, are especially important.
To investigate, researchers trained LLMs on materials science and regulatory genomics datasets with varying proportions of monofacts. They measured discovery by how often models correctly predicted scientific outputs that were not in the training data.
Higher monofact rates consistently produced more discoveries across different model families and evaluation settings. This reveals that the frequency of rarely-seen data directly controls a model's ability to innovate, offering a practical guide for data curation to enhance LLM discovery capacity.