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

⚠ This page and the underlying benchmarks contain examples of toxic and offensive stereotypes, used for the purpose of studying and mitigating bias.

Overview of deductive stereotyping and Fair-GCG reasoning-time steering — **Deductive stereotyping:** models apply population-level regularities to individual cases, producing logically coherent yet socially biased inferences. **Fair-GCG** discovers reasoning-time injection phrases that steer the model back toward fairness-aware reasoning.

Abstract

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, improve reasoning-level fairness, reduce bias in open-ended generation, and transfer to real-world fairness-sensitive tasks.

Key contributions

Deductive stereotyping. We name and characterize a dominant failure mode in fairness-sensitive reasoning, and give it a Bayesian/statistical interpretation of when and why group priors shift individual-level conclusions.
Reasoning-time steering. A mid-generation injection framework distinct from post-hoc self-reflection that nudges models toward fair inferences while they reason.
Fair-GCG. The first adaptation of GCG-style discrete prompt optimization to bias mitigation. Discovered phrases outperform existing bias-mitigation methods and transfer across models, datasets, and scales.

BibTeX

@article{deng2026fair,
  title   = {Wait, am I Being Fair? Characterizing Deductive Stereotyping and Mitigating It with Fair-GCG},
  author  = {Deng, Naihao and Zhu, Yilun and Nwatu, Joan and Scott, Clayton and Mihalcea, Rada},
  journal = {arXiv preprint},
  year    = {2026}
}