Multi-level prompting: Enhancing model performance through hierarchical instruction integration

Abstract

With the recent remarkable advancements in artificial intelligence language models, various instruction prompting techniques have been introduced across natural language processing tasks to maximize model utility and enhance performance. To address the issues of excessive generalization or over-segmentation in conventional instruction prompt design, we propose a novel framework that integrates two complementary types of instruction: granular instruction and holistic instruction. Granular instruction is an explicit prompt that provides the unique attributes of individual queries, effectively leveraging the inherent information within each query. Holistic instruction provides a structured prompt that embodies the typical characteristics of similar queries, offering a broader perspective that facilitates the extension of existing knowledge and insights. We used various pre-trained language models to validate the proposed framework to address downstream tasks that demand deep understanding and implicit knowledge. The comparative analysis demonstrated significant performance improvements. Additionally, we clearly illustrated its practical effectiveness through diverse quantitative evaluations and case studies. This study proposes a new approach to instruction prompt design, demonstrating its broad applicability to various downstream tasks and its potential to improve language model performance.