CORE 论文提出了一种基于重排器蒸馏改进多模态嵌入的组合推理方法，旨在解决MLLM嵌入模型在组合检索中区分具有相同概念但不同属性-对象绑定场景的局限性。该方法通过合成跨越五个组合匹配级别的候选列表，并引入Rank-KL目标，使嵌入模型能够重现重排器的细粒度排名。实验结果表明，该方法在多个组合推理基准测试中取得了显著的性能提升。
要点：
1. CORE方法通过重排器蒸馏技术，将重排器的组合推理能力迁移到嵌入模型中。
2. 引入Rank-KL目标，使嵌入模型能够学习到重排器的细粒度排名。
3. 在三个组合推理基准测试（COLA、SUGARCREPE++、NEGBENCH）中，CORE-RERANKER-8B取得了82.7%的总平均成绩，优于Jina-Reranker 10.7个百分点。
4. CORE-EMBED-8B在所有评估的嵌入模型中取得了最佳总平均成绩（0.666）。
5. 该方法的改进在MCMR基准测试中得到了验证，同时没有牺牲在COCO和Flickr30K上的检索性能。
MLLM-based embedding models remain limited in compositional retrieval, often failing to distinguish scenes containing the same concepts but different attribute-object bindings. Yet the same backbone can resolve such distinctions when used as a cross-attentive reranker, motivating us to distill its compositional judgments into the embedding model. We propose CORE, 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. We further introduce a graded evaluation protocol and compare contrastive learning, pairwise CoSENT, and listwise Rank-KL under the same data and tuning budget. Our comparison shows that both CoSENT and Rank-KL use the multi-level supervision more effectively than contrastive learning, with Rank-KL achieving the strongest overall performance. Across three compositional reasoning benchmarks (COLA, SUGARCREPE++, NEGBENCH), CORE-RERANKER-8B achieves an 82.7% total average, outperforming Jina-Reranker by 10.7 points, while CORE-EMBED-8B achieves the best total average (0.666) among all evaluated embedding models. The improvements transfer to the MCMR benchmark without sacrificing retrieval performance on COCO and Flickr30K.
原始发布方：HuggingFace Daily Papers（社区热门论文）
原文时间：2026-09-03 08:00:00 +08:00
