Paper claims unpaired translation between embedding spaces exposes vector database privacy risks
- The authors claim the first unsupervised method that translates text embeddings between vector spaces with no paired examples, source encoders, or predefined matches.
- The method maps embeddings into and out of a proposed universal latent representation, based on the Platonic Representation Hypothesis of shared semantic structure across models.
- Across model pairs with different architectures, parameter counts, and training datasets, the paper reports high cosine similarity after translation.
- The authors warn that an attacker with only embedding vectors may recover enough document information for classification and attribute-inference attacks against vector databases.
- The paper was first submitted to arXiv on 18 May 2025 and revised to version 4 on 26 January 2026 by Rishi Jha, Collin Zhang, Vitaly Shmatikov, and John X. Morris.
Hacker News opinions
The paper is light on technical detail and heavy on exposition. Coming from mathematics, I cannot tell whether the idea is actually novel or meaningful.
I agree it has not shown that it meets a machine learning journal's publication standard. arXiv is a preprint server, not peer review.
If pairwise-distance ordering is preserved, I would start by matching diametrically distant point pairs across the two embedding spaces. After finding enough such pairs, ideally a largest d-simplex, I would fit the scaled rigid transformation instead of solving generic graph isomorphism.
I do not buy the claim that a fully trained LLM must be pure white noise because any regularity would imply a smaller model. That assumes the trained network is already a minimal representation.
Distillation and model-printing work show that smaller parameter sets can often reproduce a network's behavior fairly well. Large models give optimization room and paths out of local minima that a lower-dimensional model may not have.
Model weights are not compressed, so interpretable redundancy exists in practice.