Vector Linking via Cross-Model Local Isometric Consistency
A recent preprint on arXiv presents a novel approach for connecting vectors from various embedding models. The research reveals that contrastive encoders, when trained independently, maintain local geometric consistency, preserving short-range distances up to a scale factor, although long-range distances may be distorted. The authors propose a geometric embedding hashing method based on iterative references, which retrieves correspondences from a limited set of paired anchors. This technique encodes each vector by measuring distances to selected paired anchors, suggests potential links through hash-space matching, and consolidates evidence using a Beta-Bernoulli posterior to establish reliable links. Tests conducted across several benchmarks and pairs of embedding models demonstrate effective and resilient linking.
Key facts
- arXiv:2605.31100
- Vector Linking recovers cross-model object correspondences using only vectors
- Independently trained contrastive encoders exhibit local geometric consistency
- Short-range distances are approximately preserved up to a scale factor
- Long-range distances are not preserved due to model-specific distortion
- Proposed method uses iterative reference-based geometric embedding hashing
- Method requires a tiny seed set of paired anchors
- Experiments across multiple benchmarks and embedding model pairs demonstrate accurate linking
Entities
Institutions
- arXiv