A Novel Long Prompt Optimization Approach to Improve Reasoning Capacity of LLMs

Large Language Models (LLMs) have demonstrated significant capabilities in solving complex open-domain tasks, often guided by comprehensive prompts. However, these prompts are typically lengthy and require substantial human effort to design. We explore the application of beam search algorithms combined with the selection and mutation operators of genetic algorithms for the automatic engineering of long prompts, aiming to enhance LLM performance in complex open-domain tasks. Our study shows that a simple beam search method, coupled with selection and mutation operations performed by large models, outperforms other methods in search efficiency. Experimental results indicate that the proposed technique achieved an average accuracy gain of 7.2% across six tasks in Big-Bench Hard, highlighting the importance of automated prompt design in fully leveraging the potential of LLMs.