Bias and Fairness in Large Language Models: Evaluation and Mitigation Techniques

Abstract: Large Language Models (LLMs) such as GPT, BERT, and LLaMA have transformed natural language processing, yet they exhibit social biases that can reinforce unfair outcomes. This paper systematically evaluates bias and fairness in LLMs across gender, race, and socioeconomic dimensions using benchmark datasets and fairness metrics. We assess bias through template-based probing, stereotype score measurement, and downstream task performance. We implement mitigation strategies including adversarial training, counterfactual data augmentation, and fairness-aware loss functions. Regression and predictive analysis reveal that token frequency and representation distance significantly correlate with bias scores. Post-mitigation analysis shows up to a 48% reduction in bias indicators with minimal accuracy trade-offs. Keywords: Large Language Models, Bias, Fairness, Evaluation, Mitigation, GPT, BERT, Counterfactual Augmentation, SHAP, LIME