The latent geometry of dream content

Dreams provide a potential unique window into human cognition, yet it remains unclear whether dream content reflects structured representational organization or largely idiosyncratic expression. Here, we analyze a large-scale dataset of dream narratives (21,607 reports) using a transformer language model to construct a high-dimensional representational space of dreams, in which each report is mapped onto a vector. We ask whether this space is structured, and if so, how different sources of information are organized within it. We find that dream content exhibits a structured latent geometry: dreams produced by different individuals, as well as by individuals of different genders, occupy systematically distinct regions of this space, particularly in more abstract semantic representations. Temporal information is also embedded in this geometry: the similarity between dreams decreases gradually as the temporal distance between them increases, revealing a continuous encoding of historical proximity, although this effect mainly reflects changes in language use and narrative conventions over time. Finally, similarity among independent descriptions of dreamers predicts similarity in their dream content, linking person-level characteristics to the organization of dream representations. Together, these results show that dreams are not random constructions but structured expressions of latent cognitive organization, in which personal, demographic and historical dimensions are systematically embedded across multiple representational levels. By making this structure explicit, the present approach provides a general computational framework for mapping the organization of internal experience.