Let's Unlearn Stereotypes Before Decision-Making: Assessing the Impact of Intrinsic Bias Mitigation on Downstream Fairness in LLMs

Research from arXiv introduces Fairness-Aware Concept Unlearning (FACU), a novel method designed to mitigate intrinsic biases in Large Language Models (LLMs) without compromising their predictive performance or language modeling quality. This development addresses a critical challenge in AI, where biased predictions can perpetuate social and economic disparities, especially as LLMs are increasingly deployed in sensitive decision-making contexts. FACU directly regularizes probability differences between stereotypical and anti-stereotypical associations, offering a targeted approach to fairness beyond general bias suppression.

Why it matters

The increasing integration of LLMs into critical decision-making processes necessitates robust mechanisms to ensure fairness and prevent the reinforcement of societal biases. Developing effective intrinsic bias mitigation techniques, such as FACU, is strategically vital for maintaining public trust in AI systems and ensuring their ethical deployment across various sectors.

Key insights

• Large Language Models (LLMs) are being used in high-stakes decision-making systems.
• Biased predictions from LLMs can exacerbate social and economic inequalities.
• A key challenge is determining if mitigating intrinsic bias directly leads to fairer downstream outcomes.
• Fairness-Aware Concept Unlearning (FACU) is introduced as a model-level mitigation technique.
• FACU adapts concept unlearning for fairness-oriented representation balancing.
• FACU regularizes probability differences between stereotypical and anti-stereotypical associations.
• The method preserves the predictive performance and language modeling quality of LLMs.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2509.16462