GTR (Generalizable T5-based dense Retriever)

Google

Structures found in this family (2)

Platonic Representation Hypothesis, Linear Subspace

By model (1)

GTR-base · 110M

Platonic Representation Hypothesis, Linear Subspace

Observations (3)

mini-vec2vec aligns embedding models with a cheap orthogonal map

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 representation intervention inverts back into gender-flipped text

Papers

mini-vec2vec: Scaling Universal Geometry Alignment with Linear Transformations (2025), Harnessing the Universal Geometry of Embeddings (2026), Intervention Lens: from Representation Surgery to String Counterfactuals (2024)