Reliability-Aware Audit of Molecular Representations for Human Olfaction

A recent preprint available on arXiv (2607.24848) conducts a reliability-focused audit of general molecular representations related to human olfaction. This research questions the prevalent belief that achieving high predictive accuracy in subsequent tasks indicates that a learned representation reliably reflects scientific structures, offers additional benefits beyond robust baselines, or effectively generalizes to unfamiliar data. The audit examines four assertions: global perceptual geometry, incremental predictive contributions beyond chemistry, replication across datasets, and the transfer of mixtures to novel components. Utilizing the Keller-Vosshall and Bierling single-molecule rating datasets alongside the Ma binary-mixture dataset, the authors assess MoLFormer and ChemBERTa against RDKit descriptors and Morgan fingerprints. The study reveals that human ratings on three attributes (intensity, pleasantness, familiarity) are reproducible across different participant groups, highlighting the necessity for stricter evaluation standards in molecular representation learning.