CORE Distills a Reranker's Compositional Judgments Into the Embedding Model, Beating Jina-Reranker by 10.7 Points

arXiv 2609.04083·medium signalXBlueskyLinkedInCopy link

MLLM embedding models fail to distinguish scenes with the same concepts but different attribute-object bindings, yet the same backbone resolves those distinctions when used as a cross-attentive reranker, which is the gap CORE exploits. It synthesizes candidate lists spanning five compositional matching levels and trains the embedding model with a listwise Rank-KL objective to reproduce the reranker's fine-grained ranking. Under matched data and tuning budget, both CoSENT and Rank-KL use multi-level supervision better than contrastive learning, with Rank-KL strongest: CORE-RERANKER-8B hits an 82.7% total average across COLA, SUGARCREPE++ and NEGBENCH, 10.7 points over Jina-Reranker, while CORE-EMBED-8B leads all evaluated embedding models at 0.666 without sacrificing COCO or Flickr30K retrieval.