AdaBoosting Improves Vision-Language Model Text Prompt Accuracy

The quality of text prompts is crucial for the classification accuracy of Vision-Language Models (VLMs). While existing few-shot prompting methods offer marginal improvements, they often fail to specifically address misclassified examples, limiting their effectiveness. This new research introduces Text Prompt Boosting (TPB), a novel framework inspired by AdaBoost, designed to overcome these limitations. TPB treats each text-prompt-based classifier as a weak learner, iteratively combining them into a robust ensemble. Its core innovation lies in explicitly focusing on hard, misclassified examples during the prompt construction process. This targeted approach allows the system to learn more effectively from limited labeled data. Extensive experiments across eleven classification benchmarks demonstrate that TPB not only boosts accuracy on the initial source model but also preserves these shot-driven gains when transferring to more powerful, larger VLMs. This cross-model transferability is a significant advantage, as previous methods often struggle to maintain improvements in such scenarios.