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 13 models from 9 families—including a decoder-based LLM (GTE-Qwen2-1.5B)—we find that similarity compression is the dominant predictor of RAT accuracy within representationally compatible families (ρ = −0.62). A second, independent factor determines whether translation is possible at all: representational compatibility, measured by RDM correlation. BERT-family encoders and newer decoder-based models share zero RDM correlation (ρ ≈ 0), creating a hard boundary that no current method—RAT, Ridge, or Procrustes—can cross. Within compatible families, RAT achieves Recall@1 from 13% to 99% across 110 directed pairs. A comparison with Procrustes, Ridge, and Affine alignment across 132 pairs reveals complementary strengths: RAT dominates with ≤50 anchor pairs (74% win rate), linear methods dominate with ≥200 (90%). Three application scenarios demonstrate practical value: same-family model upgrades retain 94% retrieval accuracy at 0.05% of re-indexing cost; lightweight query gateways achieve 98% accuracy with 6× speedup; and multi-database unified search reaches Recall@5 = 96%. 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.