[Paper Review] Disclosure and Mitigation of Gender Bias in LLMs

Xiangjue Dong, Yibo Wang+2 more|arXiv (Cornell University)|Feb 17, 2024

Business Law and Ethics5 citations

TL;DR

The paper proposes an indirect probing framework using conditional generation to reveal gender bias in LLMs without explicit gender mentions, and evaluates three mitigation strategies (Hyperparameter Tuning, Instruction Guiding, Debias Tuning) across multiple models.

ABSTRACT

Large Language Models (LLMs) can generate biased responses. Yet previous direct probing techniques contain either gender mentions or predefined gender stereotypes, which are challenging to comprehensively collect. Hence, we propose an indirect probing framework based on conditional generation. This approach aims to induce LLMs to disclose their gender bias even without explicit gender or stereotype mentions. We explore three distinct strategies to disclose explicit and implicit gender bias in LLMs. Our experiments demonstrate that all tested LLMs exhibit explicit and/or implicit gender bias, even when gender stereotypes are not present in the inputs. In addition, an increased model size or model alignment amplifies bias in most cases. Furthermore, we investigate three methods to mitigate bias in LLMs via Hyperparameter Tuning, Instruction Guiding, and Debias Tuning. Remarkably, these methods prove effective even in the absence of explicit genders or stereotypes.

Motivation & Objective

• Motivate the need to uncover gender bias in LLM outputs beyond explicit stereotypes.
• Develop an indirect probing framework that forces LLMs to reveal gender bias without explicit gender terms.
• Define metrics to quantify explicit and implicit bias and benchmark bias across multiple LLMs.
• Evaluate and compare three debiasing strategies and identify the most effective method.

Proposed method

• Introduce conditional generation probing to elicit gender bias without explicit gender mentions.
• Three probing strategies: Naturally-sourced inputs, LLM-generated inputs, and Template-based inputs containing stereotypes.
• Define explicit bias metric GAS (Gender Attribute Score) and implicit bias metrics GLD (Gender Logits Difference) and ADD (Attribute Distribution Distance).
• Probe ten LLMs across model families (LLaMA2, Vicuna, Falcon, OPT) to assess bias under each strategy.
• Evaluate three mitigation methods: Hyperparameter Tuning (temperature, Top-p, Top-K), Instruction Guiding, and Debias Tuning.
• Debias Tuning uses a QLoRA-based approach with a combined loss L = Ld + Lg + Ll to align gender distributions and reduce bias.

Experimental results

Research questions

• RQ1 RQ1: How do LLMs behave under different probing strategies when explicit stereotypes are not required?
• RQ2 RQ2: Are neutral probes more stable and suitable than Template-based prompts for bias disclosure?
• RQ3 RQ3: Do larger or alignment-trained models exhibit more or less bias, and how do mitigation strategies perform across settings.

Key findings

• All ten tested LLMs show explicit and/or implicit gender bias under at least some probing strategy.
• Template-based prompts yield the strongest bias signals, while neutral probes reveal bias more stably across models and inputs.
• Larger or aligned models tend to exhibit more explicit bias but may show reduced implicit bias (GLD/ADD) in certain conditions.
• Instruction Guiding reduces both explicit and implicit bias in many cases, especially for Naturally-sourced inputs.
• Debias Tuning consistently achieves the strongest bias mitigation across datasets and model variants, outperforming Hyperparameter Tuning and Instruction Guiding.
• Debias Tuning remains effective even when explicit genders or stereotypes are not present in the inputs.