LLMs predict material properties based on atom types and not atom positions
摘要
Nature Computational Science 在线发表的一项系统实验显示，大语言模型在预测材料性质时难以利用原子位置信息，主要依据原子类型作出判断。这一局限不会随模型规模或数据集规模增大而改变。准确预测材料性质是计算材料科学的主要挑战之一。
摘要由模型根据原文撰写，以原文为准。
正文 · 原文
Nature Computational Science, Published online: 13 October 2026; doi:10.1038/s43588-026-01044-3
Accurately predicting properties of materials is one of the main challenges in computational materials science. Systematic experiments now show that large language models (LLMs) struggle to use positional information for the prediction of those properties. This limitation does not reverse as model or dataset sizes grow.
注意转载授权 · 阅读原文
来源：Nature Computational Science · nature.com
引用原文
Nature Computational Science. LLMs predict material properties based on atom types and not atom positions[EB/OL]. (2026-10-13)[访问日期]. https://www.nature.com/articles/s43588-026-01044-3.