Text2Smell: Emergent Representations of Human Olfactory Perception in Large Language Models

Abstract: Large language models (LLMs) trained exclusively on text have recently demonstrated emergent capacities to perceive the world as humans do, suggesting that linguistic co-occurrence statistics implicitly encode aspects of human sensory experience. Yet, whether such models capture the structure of olfactory perception, one of the most complex and least understood human senses, remains unknown. In this work, we investigate whether state-of-the-art LLMs can predict human smell perception purely from linguistic cues and how their representations compare to those of molecular transformer models explicitly trained on chemical structure. We prompt LLMs to provide perceptual olfactory ratings to odorants, and evaluate their outputs against human ratings across several datasets. Surprisingly, we find that LLMs exhibit strong alignment with human perceptual judgments, comparable to, and in most cases exceeding, the performance of specialized molecular transformers. These results indicate that linguistic knowledge alone carries rich latent structure about human olfaction, bridging the gap between language and chemical perception. Our findings position LLMs as powerful linguistically grounded perceptual models and open new directions for studying sensory grounding and cross-modal representation learning through language.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Xuanjing_Huang1

Submission Number: 7870