CORE: Improving Compositional Reasoning in MLLM Embedding via Reranker Distillation
CORE is proposed, which synthesizes candidate lists spanning five compositional matching levels and introduces a Rank-KL objective that trains the embedding model to reproduce the reranker's fine-grained ranking and compares contrastive learning, pairwise CoSENT, and listwise Rank-KL under the same data and tuning budg...
Tingyu Song, Mingxin Li, Yanzhao Zhang et al.
· 0 citations