More coordinates did not make these language models better at materials prediction.
An October 12 study separates chemical composition from spatial information. Its results caution against assuming that a larger text model learns geometry from coordinate lists.
A study published in Nature Computational Science on October 12, 2026 tested how language models use category and coordinate information to predict material properties. Nawaf Alampara, Santiago Miret and Kevin Maik Jablonka report that the tested models learned category patterns more effectively than coordinate information.
What the researchers compared
The work combines controlled synthetic datasets with material-property benchmarks and a framework called MatText. In the reported tasks where structure mattered, specialized geometric architectures outperformed the language models. Increasing model or dataset scale did not resolve the coordinate limitation in these experiments.
What this result does not cover
This is evidence from particular representations, models and prediction tasks. It does not establish that every AI system is unable to reason about space, or that a current consumer chatbot has the same measured performance.
The practical question to ask
Our interpretation: when choosing a tool for a scientific prediction, ask whether the evaluation tests the information your task actually needs. Recognizing a material’s composition and using its spatial arrangement are different capabilities. A fluent explanation alone does not separate them.
Published October 14, 2026. Source paper published October 12, 2026; no experiment date is inferred. We did not reproduce its training or benchmarks.