BATprompt (By Adversarial Training prompt) is a novel method for prompt generation designed to withstand input perturbations.
It uses adversarial training techniques to generate prompts that have strong performance on perturbed tasks.
BATprompt avoids reliance on real gradients or model parameters and leverages the advanced capabilities of Large Language Models (LLMs).
Experiments show that BATprompt outperforms existing prompt generation methods, delivering robustness and performance under diverse perturbation scenarios.