Evaluating generative large language models as zero-shot semantic odor descriptor generators with applications to odor sensor prediction

Predicting odor quality from molecular structure remains a fundamental challenge in sensory science. In this study, we evaluated generative large language models (LLMs) as tools for olfactory prediction across multiple tasks. By assigning LLMs distinct cultural personas through role-play prompting, we demonstrate that odorant concentration modulates LLM-generated odor quality outputs in a zero-shot setting. This represents a form of contextual sensitivity absent in structure-based approaches. A systematic evaluation of prompt design revealed that the chemical name is the primary carrier of odor-relevant information. LLM-expressed confidence showed a modest correlation with prediction accuracy, suggesting partial utility as a reliability indicator. Surrogate modeling of LLM-generated pleasantness scores against molecular fingerprints confirmed chemically coherent structure–percept associations with substructural patterns consistent with established olfactory knowledge. Finally, feature vectors derived from LLM outputs improved the prediction of instrumental odor sensor measurements relative to conventional molecular fingerprints for specific odor categories. These findings suggest that generative LLMs can serve as a complementary approach to computational odor prediction, with particular strengths in encoding the contextual and hedonic dimensions of olfactory perception.