Wait, am I Being Fair? Characterizing Deductive Stereotyping and Mitigating It with Fair-GCG

2026-06-29

AI Summary

Large language models are prone to deductive stereotyping during reasoning—erroneously applying group-level statistical generalizations to individuals, thereby producing logically coherent yet socially biased conclusions. This work presents the first formal characterization of this phenomenon and introduces Fair-GCG, a framework that intervenes during inference by automatically searching for fairness-guided prompt phrases. Integrating adversarial phrase optimization, fairness benchmark evaluation, and statistical modeling, Fair-GCG significantly enhances unbiased reasoning across multiple models in both open-ended generation and fairness-sensitive tasks, demonstrating strong cross-task generalization.

Abstract

Warning: This paper contains several toxic and offensive statements. While reasoning generally improves fairness in recent large language models (LLMs), failures persist. In this work, we identify a failure mode, deductive stereotyping, in which models apply population-level statistical regularities to individual cases, producing logically coherent yet socially biased inferences. We provide a statistical interpretation of this phenomenon. To steer models toward fairness-aware reasoning, we propose a reasoning-time injection framework. We further introduce Fair-GCG to systematically discover effective injection phrases. Injection phrases discovered by Fair-GCG improve performance across multiple fairness benchmarks, generalize from smaller to larger LLMs, improves reasoning-level fairness, reduces bias in open-ended generation, and transfer to real-world fairness-sensitive tasks.

Problem

• deductive stereotyping
• fairness
• large language models
• social bias
• reasoning

Innovation

• deductive stereotyping
• reasoning-time injection
• Fair-GCG
• fairness-aware reasoning
• bias mitigation