# Social Bias Evaluation for Large Language Models Requires Prompt Variations

**Authors:** Rem Hida, Masahiro Kaneko, Naoaki Okazaki  
**arXiv:** [2407.03129](https://arxiv.org/abs/2407.03129)  
**Date:** 2024-07-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.03129v1/

## Key points

- Authors Rem Hida, Masahiro Kaneko, and Naoaki Okazaki discuss the relationship between large language models (LLMs) and social biases.
- They emphasize the importance of accurately evaluating and mitigating biases in LLMs.
- Previous studies have used downstream tasks as prompts to assess social biases in LLMs but often with a limited range of prompts.
- The authors conducted a comprehensive investigation into prompt variations, including task instruction, prompt examples, and debias-prompt.
- Their analysis revealed significant sensitivity to prompts leading to fluctuations in model rankings based on performance and bias evaluation.
- Tradeoffs exist between performance and social bias in LLMs due to prompt variations; reducing bias may impact performance outcomes.
- Sensitivity to prompts in advanced LLMs is attributed to ambiguity present in processed instances resulting in diverse outputs.
- Diverse prompts should be utilized when assessing social bias in LLMs for a more comprehensive understanding.

## Summary

In their paper titled "Social Bias Evaluation for Large Language Models Requires Prompt Variations," authors Rem Hida, Masahiro Kaneko, and Naoaki Okazaki delve into the intricate relationship between large language models (LLMs) and social biases. They highlight the pervasive presence of stereotypes and biases within LLMs and emphasize the importance of accurately evaluating and mitigating these biases. Previous studies have utilized downstream tasks as prompts to assess social biases in LLMs; however, these studies often employed a limited range of prompts. To address this limitation, the authors conduct a comprehensive investigation into the sensitivity of LLMs to prompt variations, including task instruction and prompt, few-shot examples, and debias-prompt. By analyzing both task performance and social bias outcomes of LLMs under different prompt settings, they uncover a significant sensitivity to prompts that leads to fluctuations in model rankings based on performance and bias evaluation. Furthermore, the study reveals tradeoffs between performance and social bias in LLMs resulting from prompt variations. Specifically, reducing bias through prompt adjustments may lead to diminished performance outcomes. The authors attribute this sensitivity to prompts in advanced LLMs to the ambiguity present in instances processed by these models, which can result in diverse outputs. Based on their experimental findings, they advocate for utilizing diverse prompts when assessing social bias in LLMs. By employing a variety of prompt settings similar to those explored in their study, researchers can gain a more comprehensive understanding of how different prompts impact social bias within LLMs.