Reproduce “Vector Linking via Cross-Model Local Isometric Consistency” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ziying Chen, Yang Cao, He Sun, Beining Yang, Tianjian Yang
Vector linking reframes cross-model embedding interoperability as partial correspondence recovery rather than global space alignment. Geometric Embedding Hashing (GEH) uses locally consistent distances to paired anchors, multi-view voting, and iterative bootstrapping to recover shared objects across black-box encoders, reaching roughly 80–88% F1 in many benchmark settings and 93.8% precision on a 5.4-million-document FEVER-scale test.
Key insights
- Contrastive encoders can preserve local distances up to an approximately uniform scale even when their global embedding spaces are not aligned; a manifold-based analysis connects this behavior to local isotropy of the induced metric.
- Multiple small and diverse anchor views are more robust than one global comparison: genuine links recur across views, while accidental matches caused by local distortion receive inconsistent support.
- Bootstrapping and multi-view aggregation are essential. Removing them reduces reported F1 to about 1.9% and 29.8%, respectively, while furthest-point sampling and distance kernels improve stability and locality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.