Glonce

New method translates embeddings across vector spaces without paired data

researchThe Glonce editorsHacker News🔥68Glonce rating68Impact70Novelty75Relevance95Surprise55Players20Credibility85

A paper posted on arXiv, first submitted on 18 May 2025 and revised through a fourth version on 26 Jan 2026, introduces what its authors call the first method for translating text embeddings from one vector space to another without any paired data, encoders, or predefined sets of matches. The approach is unsupervised: it maps any embedding to and from what the authors term a universal latent representation, a shared semantic structure that the paper ties to the Platonic Representation Hypothesis, the idea that different models trained on different data converge toward a common underlying geometry of meaning. The authors report that their translations achieve high cosine similarity across pairs of models that differ in architecture, parameter count, and training data, though the stored text gives no specific similarity score, benchmark, or named model pair.

The authors argue the technique carries a direct security consequence: because it lets someone translate an unknown embedding into a different space while preserving its geometry, an adversary who has access only to a set of embedding vectors, not the original text, can extract sensitive information about the underlying documents. They state this is enough for classification and attribute inference, meaning an attacker could infer properties of the source text without ever seeing it. The submission record names Rishi Jha as the submitter; no institution, affiliation, or full author list appears in the stored text, and no code, dataset, or concrete example of the attack being carried out is included.

Key facts

  • The paper claims the first unsupervised method to translate text embeddings between vector spaces without paired data, encoders, or predefined matches.
  • It maps embeddings to and from a universal latent representation, linked to the Platonic Representation Hypothesis of convergent model geometry.
  • The authors report high cosine similarity across model pairs differing in architecture, parameter count, and training data, without giving a specific figure.
  • The authors say the technique lets an adversary with only embedding vectors extract sensitive information sufficient for classification and attribute inference.
  • First submitted 18 May 2025 on arXiv, with a fourth revision posted 26 Jan 2026; submitter of record is Rishi Jha.

Why it matters

Vector databases and embedding-based search systems generally treat embeddings as opaque: two systems using different encoders are assumed to produce vectors that cannot be meaningfully compared or converted into each other without matched pairs of examples. This paper's central claim is that this assumption does not hold. If embeddings from unrelated models can be translated into a shared space without paired data, then embeddings are closer to a common, recoverable representation of meaning than the field has generally assumed, an idea the authors connect to the broader Platonic Representation Hypothesis about convergent model geometry.

Who it affects

The claim is aimed at anyone running a vector database or a retrieval system built on embeddings, including RAG pipelines, semantic search, and recommendation systems that store embedding vectors rather than raw text. It also concerns teams evaluating the privacy guarantees of embedding-only storage, since the paper argues that storing vectors instead of text does not by itself protect the underlying documents from inference.

How to use it

The paper is a preprint on arXiv (2505.12540), available as PDF and an experimental HTML view. The stored text does not describe the method's mechanism, name a released codebase, or specify which embedding models or datasets were tested, so there is nothing here to say how the translation technique would be applied in practice beyond the authors' own description of it as unsupervised and encoder-free.

How solid is it

The paper has gone through four revisions, from an initial submission in May 2025 to a fourth version in January 2026, indicating the authors have continued to update it over roughly eight months. It is a preprint, and the stored text does not indicate peer review or publication venue. The abstract-level claims of high cosine similarity are not accompanied by a specific quantitative score, named benchmark, or named model pairs in the available text.

Risks and caveats

No institution or affiliation is given for Rishi Jha or any co-authors, and no author list beyond the submitter appears in the stored text. The claimed cosine similarity is not given a numeric value, no specific models or datasets are named as having been tested, the translation method's internal mechanism is not described, and no real-world demonstration of the described attack on a vector database is included.