Enhancing Genetic Prompt Search Using Language Model Probability Exploitation

Abstract In the application of large pre-trained language models (PLMs), prompt-tuning techniques have demonstrated notable effectiveness in low-resource settings, such as learning from a small number of samples. Concurrently, derivative-free optimization (DFO) techniques enable prompt tuning on black-box PLMs, enhancing adaptation to specific downstream tasks. However, current DFO-based prompt-tuning methods typically rely on certain prerequisites, such as requiring additional APIs from PLMs to inject continuous prompts as hidden states or embedding vectors, or necessitating well-designed discrete manual prompts as initial states for the algorithm. To overcome these limitations and make DFO-based prompt tuning more universally applicable, we propose a novel genetic algorithm (GA). This algorithm starts with a blank prompt and drives the evolution of the prompt by leveraging prediction probabilities guided by a small-sample training set, thus selecting appropriate tokens. Experiments across multiple diverse benchmark datasets have shown that this unconditional approach significantly outperforms existing prerequisite-dependent DFO methods, such as black-box tuning, genetic prompt search, and gradient-free imperative prompt search.