Share
A new study from Johns Hopkins University suggests large language models may work differently, at least when it comes to scientific discovery. Facts that show up only once in a model’s training data appear to make it better at finding new things.
The paper is titled “Tracing the Scientific Discovery Capacity of LLMs to Their Training Data.” It was published on November 3, 2026, and it puts a name on these one-time facts: monofacts.
What the researchers actually tested
The team defines a monofact as a piece of factual information a model encounters exactly once during training.
The central claim is that these single-appearance facts meaningfully improve a model’s ability to make novel scientific discoveries.
To test this, the researchers ran experiments in two fields. The first was materials science. The second was regulatory genomics, which involved analyzing DNA sequences.
The discovery test was designed to avoid a common trap. Models were scored on whether they could find held-out materials and DNA sequences, meaning examples they had never been trained on, from after the model’s training cutoff date.
The results pointed in one direction. As monofact rates went up, discovery rates rose with them, and the researchers describe that increase as reliable.
Not just a diversity story
The authors say that explanation does not cover it. According to the study, the boost in discovery could not be attributed solely to the overall diversity of the training set.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
We respect your privacy. Unsubscribe anytime.
The team also tried to make sure the finding was not a quirk of one setup. They tested multiple open-weight model families, several sampling strategies and a range of reasoning budgets.
Open-weight models are those whose internal parameters are publicly available. That makes them easier for academic researchers to inspect and retrain than closed commercial systems.
The effect also showed some ability to carry over between scientific domains. The catch is that this transfer worked when the training data included relevant in-domain examples.
Where this research comes from
The study builds on earlier work from the same institution. In June 2026, Johns Hopkins researchers published work showing that in-context learning capabilities emerged in genomic models trained on non-language sequences such as DNA.
In-context learning is the ability of a model to pick up a task from examples given in a prompt, without being retrained.
The author list reaches beyond academia. Co-authors include figures from Microsoft and Apple, and the publication sits within a broader Johns Hopkins effort to study how empirical data and AI training interact in genomics and materials science.
What this means for how models get built
The most direct implication, as the research frames it, concerns how training data gets curated. If monofacts drive discovery, then decisions about what to keep, cut or duplicate in a dataset take on new weight.
There are limits worth keeping in view. The experiments covered two domains, materials science and regulatory genomics, and cross-domain transfer depended on having relevant in-domain examples available.
The evaluation design is also a story in itself. By scoring models on held-out examples from after the training cutoff, the researchers addressed a persistent worry in AI-for-science work: that a model’s apparent discovery is just something it quietly memorized.
