Researchers introduce the first method for translating text embeddings between vector spaces without paired data, encoders, or predefined matches. The unsupervised approach maps any embedding to and from a universal latent representation, supporting the Platonic Representation Hypothesis. Translations achieve high cosine similarity across model pairs with different architectures, parameter counts, and training datasets.

This capability raises serious security implications for vector databases. An adversary with access only to embedding vectors can extract sensitive information about underlying documents, enabling classification and attribute inference attacks. The work demonstrates that unknown embeddings can be translated into different spaces while preserving their geometry.

The research was submitted by Rishi Jha through multiple versions from May 2025 through January 2026, with the latest version v4 posted on Mon, 26 Jan 2026.