Abstract

Large Language Models (LLMs) have achieved remarkable success across diversetasks, largely driven by well-designed prompts. However, crafting and selectingsuch prompts often requires considerable human effort, significantly limitingits scalability. To mitigate this, recent studies have explored automatedprompt optimization as a promising solution. Despite these efforts, existingmethods still face critical challenges in robustness, efficiency, andgeneralization. To systematically address these challenges, we first conduct anempirical analysis to identify the limitations of current reflection-basedprompt optimization paradigm. Building on these insights, we propose 7innovative approaches inspired by traditional deep learning paradigms forprompt optimization (DLPO), seamlessly integrating these concepts intotext-based gradient optimization. Through these advancements, we progressivelytackle the aforementioned challenges and validate our methods through extensiveexperimentation. We hope our study not only provides valuable guidance forfuture research but also offers a comprehensive understanding of the challengesand potential solutions in prompt optimization. Our code is available athttps://github.com/sfasfaffa/DLPO.