Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement

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 \url{https://github.com/JP-25/KnowBias}.

Large language models can exhibit social biases that reinforce harmful stereotypes about demographic groups. Many existing debiasing methods try to suppress biased behavior directly, but they can require large amounts of bias-annotated data, generalize poorly to unseen bias types or demographics, and degrade general capabilities. We ask whether models already encode useful bias knowledge (knowledge about when a statement, ideology, or opinion is biased) even when they do not always use this knowledge in their answers. Our method, KnowBias, identifies know-bias neurons: small internal parts of the model that support its ability to recognize bias contents. KnowBias finds these neurons using only a small set of simple yes/no bias-knowledge questions, such as whether race influences problem-solving skills. Instead of retraining the model or suppressing bias-correlated neurons, KnowBias selectively enhances know-bias neurons at inference time while the model is generating an answer. Across multiple language models and social bias benchmarks, KnowBias achieves strong bias mitigation while largely preserving general capabilities, suggesting a lightweight and scalable alternative to conventional bias-suppression approaches.