A study published on 12 October in Nature Computational Science examined how large language models use positional and categorical information in scientific data. Co-authored by researchers at Friedrich Schiller University Jena and Santiago Miret of Intel Labs, it combines controlled synthetic-data experiments with materials-property benchmarks. The authors report that language models can learn composition labels such as atom types but systematically struggle with coordinates describing how atoms are arranged in space.

The gap persisted in the tested models up to 70 billion parameters and datasets of up to two million structures. Graph neural networks that explicitly model geometry outperformed language-model approaches on structure-dependent material properties. The comparison suggests language models are not suitable for every scientific prediction task; it does not report a newly produced material or laboratory validation.