# AI Study Finds More One-Off Training Examples Raise Exact Science Matches

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/50876
Published: 2026-11-07T14:31:48.000Z
Updated: 2026-11-07T14:31:48.000Z
Section: Technology

> Experiments covered materials science and gene regulation, but repeated training removed the upward trend and more one-off examples also brought more hallucinations in one test.

Language models trained with a higher share of examples seen only once produced more exact matches in materials science and gene-regulation tasks, raising questions about how researchers curate scientific training data.

The experiments used OpenAI’s gpt-oss-20b and Google’s gemma-3-12b-it, both open-weight models. Researchers affiliated with Johns Hopkins University, Microsoft and Apple varied the proportion of one-time examples, which they called “monofacts.”

In materials science, a monofact was a chemical system appearing once in the training set. In regulatory genomics, it was a DNA sequence with no other training example overlapping it by more than 150 nucleotides. The models generated material formulas to meet specified properties or DNA sequence modifications to match target regulatory activity. A result counted as a discovery only if it exactly matched a held-out target.

The team tested monofact shares from 0% to 100% in 10-percentage-point increments. In the main experiments, the discovery rate rose steadily with the share across both model families and two training stages. The pattern appeared in pass@1 evaluations and pass@16 evaluations, where a success meant at least one of 16 generated candidates matched the target.

For the materials task, training data came from a Materials Project snapshot dated Nov. 14, 2024. The researchers evaluated formulas first appearing in a Sept. 25, 2025, snapshot. The genomics task used a dataset released in October 2024, with examples held back for evaluation.

The pattern did not hold in every test. When researchers trained for three or five epochs, the upward discovery trend disappeared even though dataset diversity remained. In an in-distribution materials evaluation, a higher monofact share also increased hallucinations.

A temporally shifted materials test showed a different mix. The share of generated output classified as unobserved stayed between 98% and 99% across monofact rates, while the researchers found that this output shifted toward discoveries and away from hallucinations as the rate rose. The cross-field effect was limited, too: genomics monofacts affected materials discovery only when materials data was also included in training.

The results come from controlled experiments with two open-weight models and do not establish how commercial developers should handle rare examples in full pretraining pipelines. They also do not show that either model produced a laboratory-verified discovery.