The paper "Prompt Learning via Meta-Regularization" by Jinyoung Park, Juyeon Ko, and Hyunwoo J. Kim addresses the issue of task overfitting in prompt learning approaches for vision-language models. Pre-trained vision-language models, such as CLIP, have shown impressive performance in various computer vision tasks, but they often suffer from task-specific overfitting when fine-tuned on small datasets for specific downstream tasks. To tackle this problem, the authors propose ProMetaR (Prompt Meta-Regularization), a framework that meta-learns both the regularizer and soft prompts to improve the generalizability of prompt tuning. ProMetaR learns the regularizer and soft prompts through a bi-level optimization process, where the outer loop meta-learns the regularizer and soft prompts, and the inner loop optimizes the model parameters using the gradients of the loss and the regularizer. Additionally, ProMetaR incorporates task augmentation to generate diverse virtual tasks, which helps alleviate meta-overfitting and enhances the model's ability to generalize to new tasks. The authors provide theoretical analysis to explain how ProMetaR improves the generalizability of prompt learning from the perspective of gradient alignment. Extensive experiments on 11 image recognition datasets and four variants of ImageNet datasets demonstrate that ProMetaR significantly outperforms existing prompt learning methods in both base-to-base/base-to-new generalization and domain generalization settings. The code for ProMetaR is available at <https://github.com/mlvlab/ProMetaR>.The paper "Prompt Learning via Meta-Regularization" by Jinyoung Park, Juyeon Ko, and Hyunwoo J. Kim addresses the issue of task overfitting in prompt learning approaches for vision-language models. Pre-trained vision-language models, such as CLIP, have shown impressive performance in various computer vision tasks, but they often suffer from task-specific overfitting when fine-tuned on small datasets for specific downstream tasks. To tackle this problem, the authors propose ProMetaR (Prompt Meta-Regularization), a framework that meta-learns both the regularizer and soft prompts to improve the generalizability of prompt tuning. ProMetaR learns the regularizer and soft prompts through a bi-level optimization process, where the outer loop meta-learns the regularizer and soft prompts, and the inner loop optimizes the model parameters using the gradients of the loss and the regularizer. Additionally, ProMetaR incorporates task augmentation to generate diverse virtual tasks, which helps alleviate meta-overfitting and enhances the model's ability to generalize to new tasks. The authors provide theoretical analysis to explain how ProMetaR improves the generalizability of prompt learning from the perspective of gradient alignment. Extensive experiments on 11 image recognition datasets and four variants of ImageNet datasets demonstrate that ProMetaR significantly outperforms existing prompt learning methods in both base-to-base/base-to-new generalization and domain generalization settings. The code for ProMetaR is available at <https://github.com/mlvlab/ProMetaR>.