Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

In-Contextual Gender Bias Suppression for Large Language Models

Daisuke Oba, Masahiro Kaneko, Danushka Bollegala

TL;DR

Large Language Models encode gender biases, and traditional debiasing methods often require access to model parameters or decoding changes, unavailable for closed LLMs. The paper proposes bias suppression via textual preambles—Counterfactual (CF-*) and Descriptive (Desc-*)—inserted at inference without parameter updates, enabling user-level mitigation. Experiments on MPT-7B, OpenLLaMA-7B, and LLaMA2-7B using the Crows-Pairs benchmark show substantial bias reductions captured by Relative Bias Score, with perplexity-based preamble selection enhancing effectiveness; downstream COPA and HellaSwag performance degrades only modestly. These findings demonstrate a practical, model-agnostic approach to mitigating bias in both open and closed LLMs, with potential for multilingual extension and automatic preamble discovery.

Abstract

Despite their impressive performance in a wide range of NLP tasks, Large Language Models (LLMs) have been reported to encode worrying-levels of gender biases. Prior work has proposed debiasing methods that require human labelled examples, data augmentation and fine-tuning of LLMs, which are computationally costly. Moreover, one might not even have access to the model parameters for performing debiasing such as in the case of closed LLMs such as GPT-4. To address this challenge, we propose bias suppression that prevents biased generations of LLMs by simply providing textual preambles constructed from manually designed templates and real-world statistics, without accessing to model parameters. We show that, using CrowsPairs dataset, our textual preambles covering counterfactual statements can suppress gender biases in English LLMs such as LLaMA2. Moreover, we find that gender-neutral descriptions of gender-biased objects can also suppress their gender biases. Moreover, we show that bias suppression has acceptable adverse effect on downstream task performance with HellaSwag and COPA.

In-Contextual Gender Bias Suppression for Large Language Models

TL;DR

Large Language Models encode gender biases, and traditional debiasing methods often require access to model parameters or decoding changes, unavailable for closed LLMs. The paper proposes bias suppression via textual preambles—Counterfactual (CF-*) and Descriptive (Desc-*)—inserted at inference without parameter updates, enabling user-level mitigation. Experiments on MPT-7B, OpenLLaMA-7B, and LLaMA2-7B using the Crows-Pairs benchmark show substantial bias reductions captured by Relative Bias Score, with perplexity-based preamble selection enhancing effectiveness; downstream COPA and HellaSwag performance degrades only modestly. These findings demonstrate a practical, model-agnostic approach to mitigating bias in both open and closed LLMs, with potential for multilingual extension and automatic preamble discovery.

Abstract

Despite their impressive performance in a wide range of NLP tasks, Large Language Models (LLMs) have been reported to encode worrying-levels of gender biases. Prior work has proposed debiasing methods that require human labelled examples, data augmentation and fine-tuning of LLMs, which are computationally costly. Moreover, one might not even have access to the model parameters for performing debiasing such as in the case of closed LLMs such as GPT-4. To address this challenge, we propose bias suppression that prevents biased generations of LLMs by simply providing textual preambles constructed from manually designed templates and real-world statistics, without accessing to model parameters. We show that, using CrowsPairs dataset, our textual preambles covering counterfactual statements can suppress gender biases in English LLMs such as LLaMA2. Moreover, we find that gender-neutral descriptions of gender-biased objects can also suppress their gender biases. Moreover, we show that bias suppression has acceptable adverse effect on downstream task performance with HellaSwag and COPA.
Paper Structure (29 sections, 2 equations, 4 figures, 10 tables)

This paper contains 29 sections, 2 equations, 4 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Bias Suppression
  4. Bias Measures for Generative LMs
  5. Experiments
  6. Evaluation of Gender Bias
  7. Benchmark Dataset
  8. Bias Measures
  9. Setup for Preambles
  10. Baselines
  11. ${instruct}$:
  12. ${intervention}$:
  13. Results for Bias Suppression
  14. Downstream Evaluation
  15. Datasets and Metrics
  16. ...and 14 more sections

Figures (4)

  • Figure 1: A conceptual diagram; proposed method provides textual preambles without updating parameters, resulting in the increased likelihood of a stereotypical text (Bottom) over the usual LLMs' usage (Upper).
  • Figure 2: RBS and Accuracy-based bias scores for the three models; (Left) MPT-7B, (Center) OpenLLaMA-7B-v2, (Right) LLaMA2-7B. We use Top-$N$ preambles with lowest perplexity.
  • Figure 3: Performance drops on (Upper) COPA and (Lower) HellaSwag when using proposed preambles compared to ${nc}$, for the three models; (Left) MPT-7B, (Center) OpenLLaMA-7B-v2, (Right) LLaMA2-7B. We use Top-$N$ preambles with lowest perplexity.
  • Figure 4: RBS trends for the three models (Left) MPT-7B, (Center) OpenLLaMA-7B-v2, (Right) LLaMA2-7B, with the different number of preambles (Upper) Top-$N$ preambles with lowest perplexity, (Lower) randomly selected preambles.