PSAO: A Structured Prompt Optimisation Framework for LLMs

A new framework known as Prompt Segmentation and Annotation Optimisation (PSAO) has been developed by researchers to enhance prompts in large language models (LLMs). This approach breaks down prompts into understandable segments, such as sentences, and enriches them with human-readable tags like {not important}, {important}, and {very important}. These tags assist LLMs in determining focus and resolving ambiguities during response creation. PSAO seeks to enhance controllability and efficiency compared to current optimisation methods that deal with unstructured prompts, which can be computationally expensive and may misrepresent the original intent. The authors provide formal definitions for segmentations and annotations, showing that optimised segment-level annotations can improve LLM outputs while preserving the initial prompt. The paper is accessible on arXiv with the identifier 2605.14561.