Zero-Shot Embedding Space Translation via Relative Anchor Similarity Profiles

Different embedding models map inputs to incompatible vector spaces, forcing costly re-embedding when models are switched or combined. We present Relative Anchor Translation (RAT), a zero-shot protocol that translates between embedding spaces by comparing similarity profiles to shared anchor points, requiring no training data or learned parameters.

Evaluating RAT on 11 models from 6 families (110 directed pairs, Recall@1 13–99%), we find that similarity compression—the degree to which a model maps all inputs to a narrow angular cone—is the dominant predictor of RAT accuracy (ρ = −0.62). Z-score normalization is conditionally beneficial: it adds up to +88 points on low-compression databases but destroys accuracy (−62 points) on compressed ones. Both conditions are diagnosable from anchor embeddings alone, enabling automatic compatibility estimation before deployment. RAT also extends to cross-modal retrieval (text-only encoder searching a vision space at 21.2% Recall@1, 106× random baseline, zero visual training).

Code and data: https://github.com/jiro-prog/rat-experiment

Implementation was conducted using Claude Code (Anthropic). Experimental design, analysis, and manuscript drafting were assisted by Claude (Anthropic). The author takes full responsibility for all content.