Zhenzhen Hu

We introduce GARE, a gap-aware retrieval framework that learns pair-specific increments to alleviate optimization tension and false-negative noise in cross-modal alignment, achieving better uniformity and semantic structure.