Language Organizes Color Geometry in Multilingual Encoders

We use Representational Similarity Analysis (RSA) to analyze whether the relational geometry of color representations is more stable across multilingual encoders within a language or across languages within an encoder. Across six languages and five multilingual encoders, we construct one representational dissimilarity matrix (RDM) from Berlin and Kay's (1969) eleven basic color terms for each of the 30 language--encoder pairs. We find that RDMs of the same language correlate at a mean Spearman $\rho = .495$ across encoders, compared with $\rho = .231$ for different languages within the same encoder. The effect survives single-term omission and disjoint context resampling. Across the models and corpora studied here, relational color geometry therefore carries a stable language-conditioned signal that generalizes across encoder architectures.