Reconfirm Instruction Prompting for LLM in In-context Learning

Abstract

There has been a recent rapid surge of interest in research on prompt techniques for large language models (LLMs). Few-shot learning (FSL), also referred to as in-context learning, is the most well-known prompting technique that appends several task examples to a prompt. In this letter, we introduce a simple, yet effective prompting to improve the capabilities of in-context learning with LLMs. We evaluate our prompting across six datasets consisting of standard English and Japanese language tasks using several LLMs to validate its effectiveness and generality.