CORE: Improving Compositional Reasoning in MLLM Embedding via Reranker Distillation

Tingyu Song, Mingxin Li, Yanzhao Zhang, Dingkun Long, Chu Liu et al.|September 3, 2026arXiv

Key Takeaway

Distilling ranking knowledge from a reranker into embedding models significantly improves compositional reasoning—the ability to distinguish scenes by attribute-object relationships—achieving 82.7% on compositional benchmarks while maintaining standard retrieval performance.

Summary

This paper improves how multimodal AI models understand complex visual scenes by teaching embedding models to better distinguish between scenes with the same objects but different arrangements.

multimodal training evaluation

Key Terms

multimodal-embedding compositional-reasoning reranker rank-kl knowledge-distillation