Vision-language models (VLMs) have achieved remarkable progress in tasks involving both language and vision. However, the distribution gap between training and testing data often degrades their performance during inference. To address this challenge, test-time adaptation (TTA) methods have been introduced. While existing TTA methods have demonstrated potential, they require significant computational resources or fail to effectively address the distribution gap, limiting their practicality in real-world scenarios. To overcome these limitations, we propose a novel framework, test-time adaptation via distribution-aware guidance (TAG). TAG incorporates two key components: distribution-aware guidance and uncertainty-driven ensemble. The distribution-aware guidance module approximates the testing data distribution and generate auxiliary predictions. The uncertainty-driven ensemble dynamically adjusts the contribution of guidance based on the uncertainty of the model’s original predictions. Extensive experiments on multiple benchmarks demonstrate that TAG achieves state-of-the-art performance while maintaining high computational efficiency.