Tracing the Scientific Discovery Capacity of LLMs to Their Training Data

The paper investigates how the scientific discovery capability of large language models (LLMs) can be traced back to their training data. By focusing on ‘monofacts’—facts that appear only once in the training set—the authors develop a theoretical framework linking monofact rates to discovery potential. Experiments in materials science and regulatory genomics show that increasing monofact rates consistently boosts discoveries across different model families, sampling strategies, and reasoning budgets, and that this effect is not solely due to training‑set diversity and can transfer across domains when paired with in‑domain data.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Computation and Language.

• llms