Adaptive Generation of Bias-Eliciting Questions for LLMs

2025-09-21 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Researchers from ETH Zurich introduced a novel counterfactual bias evaluation framework that automatically generates realistic, open-ended questions to elicit biased behavior in large language models (LLMs). This framework iteratively mutates and selects bias-inducing questions over sensitive attributes like sex, race, or religion, using an LLM-based judge to score responses across dimensions including bias strength, relevance, acknowledgment, and refusal. Leveraging this framework, they constructed CAB, a human-verified benchmark comprising 405 questions. An analysis of nine state-of-the-art LLMs on CAB revealed that GPT-5 and Claude-4-Sonnet exhibit the lowest levels of bias, while Grok-4 and Qwen3-235B show the highest. The study also found that implicit questions, using stereotypical traits, reduced observed bias by approximately 40% compared to explicit questions.

Key takeaway

For AI Scientists and Machine Learning Engineers developing or deploying LLMs, this research highlights the inadequacy of traditional bias benchmarks. You should integrate adaptive, counterfactual bias evaluation frameworks like CAB into your model development lifecycle to uncover subtle, real-world biases across sensitive attributes. Prioritize models like GPT-5 or Claude-4-Sonnet, which show lower bias, and be aware that implicit prompts can still elicit harmful biases.

Key insights

Adaptive, counterfactual question generation reveals nuanced LLM biases beyond simplistic templated benchmarks.

Principles

• Iterative question mutation optimizes bias elicitation.
• Multi-dimensional scoring captures nuanced bias types.
• Counterfactual comparisons are key for accurate bias detection.

Method

A genetic algorithm iteratively refines open-ended questions using an LLM generator, a target LLM for responses, and an LLM judge scoring bias, relevance, acknowledgment, and refusal to maximize a fitness score.

In practice

• Use CAB to benchmark LLMs for sex, race, and religion biases.
• Develop adaptive question generation for specific bias domains.
• Distinguish between explicit bias and bias acknowledgment.

Topics

• LLM Bias Evaluation
• Counterfactual Question Generation
• CAB Benchmark
• Algorithmic Fairness
• Generative AI Safety
• GPT-5

Code references

• eth-sri/cab

Best for: AI Architect, Research Scientist, CTO, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Critic-Free, Not Bias-Free: Correcting Advantage Bias in RL from Verifier Feedback — Group-based RLHF methods for LLMs exhibit a bias that hinders learning on difficult prompts.
• One More Turn, Less Regret: A Regret-Based Multi-Turn Benchmark for LLMs' Clarification Policies — LLM clarification is a sequential decision problem requiring optimal timing and stopping, not just isolated question quality.
• True Positive Weekly #166 — The issue highlights diverse AI advancements, from genomic challenges to novel LLM behaviors and practical tools.
• Prompt Perturbations Reveal Human-Like Biases in Large Language Model Survey Responses — LLMs used in surveys exhibit human-like biases, especially recency bias, necessitating robust prompt design and testing.
• When Should LLMs Search? Counterfactual Supervision for Search Routing — Counterfactual outcomes from no-search and forced-search traces provide a model-specific oracle for LLM search routing.
• Preconditioned Test-Time Adaptation for Out-of-Distribution Debiasing in Narrative Generation — CAP-TTA dynamically debiases LLMs against emergent biases via triggered, preconditioned test-time adaptation, improving safety and fluency.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.