Generalizable Meta Prompt Learning for Vision-Language Model

Vision-language models (VLMs) are highly sensitive to prompt phrasing, which can substantially affect performance. While automatic prompt tuning methods improve performance in low-data scenarios, these methods still rely on handcrafted prompt initialization, leaving room for further performance improvement. To address this, we propose a lightweight meta-prompt learning framework that uses Model-Agnostic Meta-Learning (MAML) combined with gradient clipping to train generalizable prompts under simulated few-shot settings. Unlike prior approaches that require large-scale pretraining and introduce auxiliary networks, our method preserves the original prompt tuning structure, offering better interpretability and practical applicability. Through experiments on 11 diverse datasets, we show that our meta-prompts consistently outperform handcrafted prompts in zero-shot classification and, when used for fine-tuning, maintain or even improve generalization performance. Our findings highlight the potential of simple, data-efficient meta-prompt strategies to improve robustness and transferability in prompt tuning.