# Harnessing the Universal Geometry of Embeddings

**Authors:** Rishi Jha, Collin Zhang, Vitaly Shmatikov, John X. Morris  
**arXiv:** [2505.12540](https://arxiv.org/abs/2505.12540)  
**Date:** 2025-05-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.12540v2/

## Key points

- Groundbreaking method for translating text embeddings between different vector spaces without the need for paired data, encoders, or predefined matches
- Leveraging a universal latent representation based on the Platonic Representation Hypothesis
- Achieved high cosine similarity across various model pairs with differing architectures, parameter counts, and training datasets
- Ability to translate unknown embeddings into a different space while preserving their geometric properties
- Implications for security of vector databases by preventing adversaries from extracting sensitive information from embedding vectors alone
- Examples of inversions inferring entities and content from email data
- Stands out by learning how to translate representations across spaces without any paired data
- Opens up new possibilities for translating text embeddings and enhancing security measures in vector databases
- Contributes to existing research on representation alignment, unsupervised transport, and bridging modality gaps in neural networks
- Potential advancements in natural language processing and multi-modal integration through further exploration

## Summary

The paper introduces a groundbreaking method for translating text embeddings between different vector spaces without the need for paired data, encoders, or predefined matches. This unsupervised approach leverages a universal latent representation that is hypothesized to be a universal semantic structure according to the Platonic Representation Hypothesis. The translations achieved high cosine similarity across various model pairs with differing architectures, parameter counts, and training datasets. The significance of this method lies in its ability to translate unknown embeddings into a different space while preserving their geometric properties. This has profound implications for the security of vector databases as adversaries could potentially extract sensitive information from embedding vectors alone, enabling classification and attribute inference. The paper provides examples of inversions that infer entities and content from email data. It also discusses related work in representation alignment, unsupervised transport, embedding inversion, and bridging modality gaps. The method presented in the paper stands out by not only measuring similarity between representations but also learning how to translate them across spaces without any paired data. Overall,this innovative approach opens up new possibilities for translating text embeddings and has the potential to enhance security measures in vector databases. The paper contributes to existing research on representation alignment and unsupervised transport while offering a unique perspective on bridging modality gaps in neural networks. Further exploration of this method could lead to advancements in natural language processing and multi-modal integration.