Raw Data Matters: Enhancing Prompt Tuning by Internal Augmentation on Vision-Language Models

On November 13, 2025, a significant advancement in AI was published in arXiv, detailing a novel approach known as Augmentation-driven Prompt Tuning (AugPT). This method addresses the limitations of existing data amplification strategies that depend heavily on external knowledge sources, which can be costly and inefficient. By focusing on self-supervised augmentation of unlabeled images, AugPT enhances the fine-tuning process of CLIP-based models. The introduction of a unique gating mechanism based on consensus testing allows for the filtering of noisy samples, thereby improving the quality of the augmented data. Extensive experiments have validated that AugPT not only boosts model performance but also enhances generalization capabilities, all without the need for additional external knowledge. This innovation represents a pivotal shift in how vision-language models can be trained more effectively and economically.