Adaptive Generation of Bias-Eliciting Questions for LLMs

TL;DR

The paper presents a counterfactual bias evaluation framework that automatically generates realistic, open-ended questions to elicit biased behavior from LLMs. By combining an adaptive generation loop with a multi-dimensional judge, it produces CAB, a human-verified benchmark for cross-model bias analysis across sex, race, and religion. The approach demonstrates that frontier models like GPT-5 exhibit comparatively lower bias but still reveal meaningful, domain-specific biases, underscoring the need for ongoing mitigation. The work emphasizes reproducibility, broad domain coverage, and the value of implicit versus explicit question framing in bias detection, offering a scalable path toward fairer AI systems.

Abstract

Large language models (LLMs) are now widely deployed in user-facing applications, reaching hundreds of millions worldwide. As they become integrated into everyday tasks, growing reliance on their outputs raises significant concerns. In particular, users may unknowingly be exposed to model-inherent biases that systematically disadvantage or stereotype certain groups. However, existing bias benchmarks continue to rely on templated prompts or restrictive multiple-choice questions that are suggestive, simplistic, and fail to capture the complexity of real-world user interactions. In this work, we address this gap by introducing a counterfactual bias evaluation framework that automatically generates realistic, open-ended questions over sensitive attributes such as sex, race, or religion. By iteratively mutating and selecting bias-inducing questions, our approach systematically explores areas where models are most susceptible to biased behavior. Beyond detecting harmful biases, we also capture distinct response dimensions that are increasingly relevant in user interactions, such as asymmetric refusals and explicit acknowledgment of bias. Leveraging our framework, we construct CAB, a human-verified benchmark spanning diverse topics, designed to enable cross-model comparisons. Using CAB, we analyze a range of LLMs across multiple bias dimensions, revealing nuanced insights into how different models manifest bias. For instance, while GPT-5 outperforms other models, it nonetheless exhibits persistent biases in specific scenarios. These findings underscore the need for continual improvements to ensure fair model behavior.

Figures (17)

• Figure 1: An overview of our approach. Our core framework iteratively generates and refines bias-inducing questions for a given attribute and model to increase fitness. The fitness of each question is determined by an LLM-based evaluator, which scores bias in model answers along multiple dimensions. By applying this framework to multiple attributes and models, we construct CAB, a human-verified benchmark of bias-inducing questions for systematic comparison of LLMs.
• Figure 2: Sample questions from CAB along with biased responses from frontier LLMs. We provide several full CAB samples including prompts, bias scores, and responses in \ref{['app:examples']}.
• Figure 3: Distribution of questions in CAB across superdomains, grouped by sensitive attribute.
• Figure 4: Average fitness of frontier models across all attributes on explicit questions in CAB.
• Figure 5: Avg. fitness (drop) of frontier models across all attributes on implicit questions in CAB.