CORE distills a reranker's compositional judgments into the embedding model, beating Jina-Reranker by 10.7 points.

arXiv 2609.04083 exploits a gap I hadn't seen named: MLLM embedding models can't distinguish scenes with the same concepts but different attribute-object bindings, while the same backbone resolves those distinctions when run as a cross-attentive reranker. Synthesizing candidate lists across five compositional matching levels and training with a listwise Rank-KL objective gets CORE-RERANKER-8B to 82.7% average across COLA, SUGARCREPE++ and NEGBENCH, with COCO and Flickr30K retrieval intact.