Less can be more for predicting properties with large language models.

Large language models (LLMs) up to 70 billion parameters consistently fail to learn coordinate information in coordinate-category datasets, despite excelling at category patterns.

• Why it matters: Accurate property prediction in materials science often depends on structural (coordinate) information, yet current LLMs may lack this capability, limiting their usefulness for such tasks.
• What they did: Using synthetic datasets with tunable coordinate and category contributions and a benchmarking framework (MatText), the study systematically evaluated LLMs’ ability to learn from coordinate data.
• The result: Specialized geometric architectures, such as Graph Neural Networks, outperform LLMs in property prediction tasks dominated by structural effects, establishing a 'GNN-LM wall' in performance benchmarks.