Let’s Unlearn Stereotypes Before Decision-Making: The Impact of Intrinsic Bias Mitigation on Downstream Fairness in LLMs

Abstract

Large Language Models (LLMs) are increasingly integrated into high-stakes decision-making systems, where biased predictions can amplify existing social and economic disparities. Although prior work has studied intrinsic representational biases in LLMs and unfair downstream behavior, it remains unclear whether mitigating intrinsic bias leads to fairer outcomes in real-world decision-making tasks. A key challenge is that intrinsic and extrinsic mitigation strategies are typically studied independently and intervene at different stages of the pipeline, limiting our understanding of how representational debiasing propagates to downstream fairness outcomes. Unlike prior work focused primarily on correlation analyses, our study directly evaluates whether intrinsic debiasing interventions translate into measurable downstream fairness improvements. For this purpose, we propose Fairness-Aware Concept Unlearning (FACU), an intrinsic bias mitigation method that adapts concept unlearning, originally developed for safety and privacy objectives, to fairness-oriented bias mitigation. Unlike suppression-based unlearning approaches, FACU explicitly regularizes probability differences between stereotypical and anti-stereotypical associations, encouraging more balanced demographic associations while preserving predictive performance and language modeling quality. We evaluate FACU across three open-source LLMs, multiple intrinsic bias benchmarks, and several downstream benchmark datasets derived from socio-economic decision-making tasks. Experiments consider both LLM-as-feature-extractor and fine-tuned classifier deployments. Our results show that FACU produces statistically significant reductions in intrinsic bias that are associated with downstream fairness improvements across most settings, multiple datasets, models, and fairness metrics, without significantly degrading predictive performance. Furthermore, combining FACU with extrinsic mitigation strategies such as counterfactual data augmentation and self-debiasing yields further fairness improvements. Together, our findings suggest that fairness-aware intrinsic bias mitigation can improve downstream fairness in LLM-based socio-economic decision-making systems. These results highlight the importance of shared mitigation efforts across both model development and downstream deployment stages.