LLM-Informed Discrete Prompt Optimization

Zeeshan Memon, Muhammad Arham, Adnan Ul-Hasan, Faisal Shafait

Keywords: Prompt Optimization, Large Language Models, Prompt Engineering

Abstract: The advent of Large Language Models (LLMs) has significantly improved NLP tasks, but their performance depends on effective prompt engineering, where engineers iteratively craft prompts by observing the dynamics of LLMs. With the rising number of LLMs, each trained on different data sources and thus exhibiting different internal sensitivities, prompt engineering has become an increasingly cumbersome task. The solution to these challenges lies in an automated and reliable model capable of suggesting optimized prompts and adapting to various LLMs. Previous works have primarily focused on training learnable vectors or identifying discrete prompts, which were effective for earlier, smaller language models. However, contemporary LLMs require coherent text prompts tailored to their specific training instructions. In this paper, we address this gap by proposing a methodology for training a lightweight model that not only produces legible, optimized prompts but also adapts to different LLMs. The proposed methodology has demonstrated significant performance improvements with optimized prompts across different LLMs.

Submission Number: 68