KnowBias: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement

TL;DR

This paper tackles social bias in large language models by proposing KnowBias, which strengthens bias-knowledge representations rather than suppressing biased outputs. It identifies know-bias neurons through attribution-based analysis guided by a compact set of bias-knowledge questions and enhances these neurons at inference time via multiplicative scaling, with no retraining. Empirical results show state-of-the-art debiasing across multiple datasets and backbones while preserving general language and reasoning capabilities, and the method generalizes across bias types and demographics with high data efficiency (requiring only $q=45$ questions). The approach offers a scalable, robust alternative to traditional suppression-based debiasing, with practical deployment advantages and potential extensions to broader normative objectives.

Abstract

Large language models (LLMs) exhibit social biases that reinforce harmful stereotypes, limiting their safe deployment. Most existing debiasing methods adopt a suppressive paradigm by modifying parameters, prompts, or neurons associated with biased behavior; however, such approaches are often brittle, weakly generalizable, data-inefficient, and prone to degrading general capability. We propose \textbf{KnowBias}, a lightweight and conceptually distinct framework that mitigates bias by strengthening, rather than suppressing, neurons encoding bias-knowledge. KnowBias identifies neurons encoding bias knowledge using a small set of bias-knowledge questions via attribution-based analysis, and selectively enhances them at inference time. This design enables strong debiasing while preserving general capabilities, generalizes across bias types and demographics, and is highly data efficient, requiring only a handful of simple yes/no questions and no retraining. Experiments across multiple benchmarks and LLMs demonstrate consistent state-of-the-art debiasing performance with minimal utility degradation. Data and code are available at https://github.com/JP-25/KnowBias.

Figures (4)

• Figure 1: Large Language Models (LLMs) show social bias but also know this bias simultaneously. Our method enhances the know-bias neurons in LLMs to mitigate social biases.
• Figure 2: KnowBias workflow (see Algorithm \ref{['alg:knowbias']} for details about know-bias neurons identifications and enhancement).
• Figure 3: Effect of the number of bias-knowledge questions on debiasing performance. ICAT score ($\uparrow$) for Stereoset-Inter.
• Figure 4: Hyperparameter study for KnowBias on Llama-3.2-3B-Instruct. ICAT score ($\uparrow$) for Stereoset-Inter.