Study reveals significant limitations of AI language models in biomolecular modeling

Researchers created BioMol-LLM-Bench, a comprehensive benchmark testing 13 large language models across 26 biomolecular tasks, finding that current LLMs excel at classification but struggle with regression problems and that chain-of-thought prompting may actually hurt performance. The findings suggest hybrid architectures and targeted fine-tuning approaches could improve AI's role in molecular discovery.