MATH · IN · MODELS

GTE (General Text Embeddings)

Alibaba

Structures found in this family (2)

Platonic Representation Hypothesis, Linear Direction

By model (2)

GTE-base-en-v1.5 · 109M
Platonic Representation Hypothesis
GTE-small · 33M
Linear Direction

Observations (2)

vec2vec learns a fully unpaired translation between six independently-trained text embedding models (gtr, clip, e5, gte, stella, granite), recovering cosine similarity up to 0.92 and top-1 retrieval accuracy up to 100% with no paired training data, and an ablation shows the explicit Gram-matrix-preservation (VSP) loss is causally necessary -- removing it alone collapses top-1 accuracy from 0.91 to 0.00
A von Mises-Fisher grounding direction separates grounded from hallucinated responses

Papers

Harnessing the Universal Geometry of Embeddings (2026), A Geometric Taxonomy of Hallucinations in Large Language Models (2026)