Don't Use LLMs to Make Relevance Judgments

Abstract

Relevance judgments and other truth data for information retrieval (IR) evaluations are created manually. There is a strong temptation to use large language models (LLMs) as proxies for human judges. However, letting the LLM write your truth data handicaps the evaluation by setting that LLM as a ceiling on performance. There are ways to use LLMs in the relevance assessment process, but just generating relevance judgments with a prompt isn’t one of them.