Tracing the Scientific Discovery Capacity of LLMs to Their Training Data
Facts seen just once during training increased scientific discoveries in the authors’ experiments.
The researchers manipulated the rate of these “monofacts” in training data for materials science and regulatory genomics. Higher rates consistently produced more discoveries across two open-weight model families, sampling strategies and reasoning budgets. Training-set diversity alone did not explain the relationship. The effect also transferred across scientific domains when paired with in-domain data.
ArXiv · AI/CL/LG's note
The researchers manipulated the rate of these “monofacts” in training data for materials science and regulatory genomics. Higher rates consistently produced more discoveries across two open-weight model families, sampling strategies and reasoning budgets. Training-set diversity alone did not explain the relationship. The effect also transferred across scientific domains when paired with in-domain data.
ArXiv · AI/CL/LG's note
score 5