Anchor Divergences: specifying context‑specific semantic geometry for contrastive representations
The new arXiv preprint (arXiv:2610.06919v1) introduces "Anchor Divergences," a framework linking probability distributions over anchors to Bregman geometries on fixed contrastive representation spaces. The paper shows how modeling anchor distributions yields a family of context‑specific similarity geometries and reports retrieval experiments demonstrating that Anchor Divergences can specify semantic similarity efficiently and effectively.