Cross-modal retrieval is crucial for discovering latent correspondences across different modalities. However, existing methods typically assume that training data are well-aligned, an unrealistic assumption since real-world datasets inevitably contain noisy correspondences. Many current approaches attempt to handle noise using strategies borrowed from single-modal classification, such as the small-loss trick, to identify clean training pairs. However, our experiments reveal that such small-loss-based strategies are less effective for multi-modal tasks due to the inherent modality gaps. Through comprehensive analysis, we observe that the deviation directions between paired image-caption features, termed Sample-level Alignment Drift (SAD), are compact and data-dependent. Leveraging this discovery, we introduce the Modality Gap Corrected Similarity (MGCS) framework that can more accurately measure the semantic distances of cross-modal samples, dynamically compensating for misalignments. Within MGCS, we can achieve more reliable noisy data separation to promote correct supervision during cross-modal matching model training. Extensive experiments on three widely used noisy correspondence benchmarks demonstrate that MGCS significantly surpasses current state-of-the-art methods.

In-context learning (ICL), a paradigm derived from large language models (LLMs), holds significant promise but is notably sensitive to the choice of input demonstrations. While numerous methodologies have been developed to select the optimal demonstrations from existing datasets, our work alternatively proposes to generate representative demonstrations through a Distillation-based Demonstration Generation (DDG) framework. Specifically, our approach aims to generate demonstrations that encapsulate the essential attributes of the target dataset. Rather than optimizing these demonstrations directly, we design a generative model and try to refine it by minimizing the discrepancies between the calculative models trained on generated demonstrations and the original datasets respectively. Additionally, we leverage a teacher-student framework to stabilize the training process and improve the quality of the synthesized samples. Extensive experiments conducted across ten prevalent text datasets demonstrate that our DDG method substantially outperforms existing state-of-the-art methodologies. Our code will be available at https://github.com/wwyq1/DDG.