Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Group Fairness Lens for Large Language Models

Guanqun Bi, Yuqiang Xie, Lei Shen, Yanan Cao

TL;DR

The paper introduces a group fairness lens for evaluating LLM bias using a hierarchical dimension-target-attribute schema and the GFair dataset drawn from real-world sources. It pairs this with a novel statement organization task to surface subtle biases, and proposes GF-Think, a chain-of-thought–based mitigation approach. Experiments across open-source and commercial LLMs reveal substantial cross-model variability in social bias and group fairness, highlighting safety concerns and the need for robust, dimension-aware debiasing. The work demonstrates significant bias reduction through GF-Think and provides a foundation for broader, multilingual fairness evaluation in large language models.

Abstract

The need to assess LLMs for bias and fairness is critical, with current evaluations often being narrow, missing a broad categorical view. In this paper, we propose evaluating the bias and fairness of LLMs from a group fairness lens using a novel hierarchical schema characterizing diverse social groups. Specifically, we construct a dataset, GFAIR, encapsulating target-attribute combinations across multiple dimensions. Moreover, we introduce statement organization, a new open-ended text generation task, to uncover complex biases in LLMs. Extensive evaluations of popular LLMs reveal inherent safety concerns. To mitigate the biases of LLMs from a group fairness perspective, we pioneer a novel chainof-thought method GF-THINK to mitigate biases of LLMs from a group fairness perspective. Experimental results demonstrate its efficacy in mitigating bias and achieving fairness in LLMs. Our dataset and codes are available at https://github.com/surika/Group-Fairness-LLMs.

A Group Fairness Lens for Large Language Models

TL;DR

The paper introduces a group fairness lens for evaluating LLM bias using a hierarchical dimension-target-attribute schema and the GFair dataset drawn from real-world sources. It pairs this with a novel statement organization task to surface subtle biases, and proposes GF-Think, a chain-of-thought–based mitigation approach. Experiments across open-source and commercial LLMs reveal substantial cross-model variability in social bias and group fairness, highlighting safety concerns and the need for robust, dimension-aware debiasing. The work demonstrates significant bias reduction through GF-Think and provides a foundation for broader, multilingual fairness evaluation in large language models.

Abstract

The need to assess LLMs for bias and fairness is critical, with current evaluations often being narrow, missing a broad categorical view. In this paper, we propose evaluating the bias and fairness of LLMs from a group fairness lens using a novel hierarchical schema characterizing diverse social groups. Specifically, we construct a dataset, GFAIR, encapsulating target-attribute combinations across multiple dimensions. Moreover, we introduce statement organization, a new open-ended text generation task, to uncover complex biases in LLMs. Extensive evaluations of popular LLMs reveal inherent safety concerns. To mitigate the biases of LLMs from a group fairness perspective, we pioneer a novel chainof-thought method GF-THINK to mitigate biases of LLMs from a group fairness perspective. Experimental results demonstrate its efficacy in mitigating bias and achieving fairness in LLMs. Our dataset and codes are available at https://github.com/surika/Group-Fairness-LLMs.
Paper Structure (33 sections, 4 equations, 8 figures, 14 tables, 1 algorithm)

This paper contains 33 sections, 4 equations, 8 figures, 14 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Evaluating bias and fairness in LLMs
  4. Mitigating Biases in LLMs
  5. Problem Formulation
  6. Dataset Construction
  7. Schema Definition
  8. Pipeline of Data Collection
  9. Evaluation Methodology
  10. Experiments
  11. Models
  12. Evaluation Metrics
  13. Social Bias.
  14. Group Fairness.
  15. Implementation Details
  16. ...and 18 more sections

Figures (8)

  • Figure 1: Some examples that lack group fairness. For the same attribute with only the target altered, the output shows toxicity towards the target middle-aged but was safe for the target elderly. Additionally, when shifting the dimension from age to nationality, the LLM declines to comment.
  • Figure 2: Relation between bias and fairness.
  • Figure 3: An illustration of the statement organization evaluation method.
  • Figure 4: The significant difference results of the GPT-4 model across dimensions. Darker shades indicate lower p-values. p < 0.05 cells with black-blue color indicates a statistically significant difference between the compared groups.
  • Figure 5: Standard deviation between targets under each dimension.
  • ...and 3 more figures