#hard negative sampling

UniME introduces a two-stage framework that significantly improves multimodal representation learning by leveraging textual knowledge distillation and hard negative instruction tuning, outperforming existing models on multiple benchmarks.