Embedding Models Measure in Peculiar Ways

Embedding spaces fail to consistently represent objective physical dimensions like mass, distance, and volume. Analysis shows that similarity scores for physical measurements are driven primarily by superficial string patterns rather than semantic magnitude. Recalibrating embedding distances does not significantly improve alignment with actual physical properties.