# Humans or LLMs as the Judge? A Study on Judgement Biases

**Authors:** Guiming Hardy Chen, Shunian Chen, Ziche Liu, Feng Jiang, Benyou Wang  
**arXiv:** [2402.10669](https://arxiv.org/abs/2402.10669)  
**Date:** 2024-02-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.10669v1/

## Key points

- The study developed a novel framework to investigate biases in human and Large Language Model (LLM) judges.
- Five specific types of biases were focused on, with significant inclinations found in both human and LLM judges.
- Weaknesses identified can be exploited under LLM judgment.
- Importance of shedding light on vulnerabilities and biases in both human and LLM judges when evaluating performance for deeper understanding.
- Emphasis on developing robust evaluation systems and encouraging expanded evaluation efforts using an open-sourced dataset for further research.

## Summary

The authors of this study developed a novel framework to investigate biases in human and Large Language Model (LLM) judges. They focused on five specific types of biases and found significant inclinations in both human and LLM judges. The study also revealed that these weaknesses can be exploited under LLM judgment. By shedding light on the vulnerabilities and biases present in both human and LLM judges when evaluating performance, the authors hope to contribute to a deeper understanding of these issues. They emphasize the importance of developing robust evaluation systems and encourage expanded evaluation efforts by leveraging an open-sourced dataset for further research. <br>
<br>
 The study involved 60 human evaluators from various countries who were selected based on specific criteria related to English proficiency, logic skills, and educational background. These evaluators participated in experiments where they were tasked with evaluating the quality of answers to questions. The control group performed 2162 evaluation tasks while the experiment group performed 2367 tasks. Detailed instructions were provided to ensure a clear understanding of the evaluation criteria.<br>
<br>
 Evaluators were instructed to focus solely on the semantic quality of answers and disregard non-semantic factors such as tone or format. They were also given guidelines on how to handle situations where both responses seemed equally good or bad.<br>
<br>
 Overall, this study highlights the presence of biases in both human and LLM judges when evaluating performance. It emphasizes the need for robust evaluation systems and encourages further research using an open-sourced dataset.<br>
<br>
 Through their research, the authors aim to contribute towards a deeper understanding of biases inherent in both LLMs and human evaluators. By encouraging expanded evaluation efforts, they hope to enhance understanding of bias in LLMs and human judgment processes.