arXiv:2610.12781v1 — "Distilled Optimal Transport" for representation alignment from a few paired anchors
The new arXiv preprint (2610.12781v1) proposes 'Distilled Optimal Transport', a method that represents embeddings by their similarities to a small set (N≈100–500) of paired anchors, aligns the two modalities by solving a single entropic optimal transport problem with marginal constraints to reduce hubness, and distills the transport plan into lightweight encoders so the mapping is inductive. The paper reports improved image–text and molecule–text alignment versus prior methods at the same anchor budget on Flickr30k, MS-COCO and ChEBI-20.