CORE: Distilling Compositional Reasoning into MLLM Embeddings
September 2, 2026
CORE improves MLLM embedding compositional reasoning by distilling cross-attentive reranker judgments using a Rank-KL objective. It targets the ability to distinguish scenes with identical concepts but different attribute-object bindings.
HOW THIS AFFECTS YOU
●
builderYou can improve retrieval accuracy for complex queries where object-attribute relationships matter.
●
researcherThe Rank-KL objective provides a way to transfer fine-grained reasoning from rerankers to embedding models.