FairStress Benchmark Evaluates LLM Demographic Fairness Beyond Final Decisions

Introducing FairStress

A new benchmark called FairStress has been released to evaluate demographic fairness in large language models (LLMs). This tool goes beyond simply checking if a model makes the same final decision when demographic identities are swapped. It focuses on how readily a model's decision changes when presented with persuasive arguments.

Addressing Hidden Asymmetries

Standard fairness benchmarks often test if a model selects the same candidate in scenarios where only demographic identities differ. However, a consistent final answer can mask underlying biases in how a model reacts to challenges. FairStress is designed to uncover these 'hidden asymmetries' by observing if a model is more or less likely to alter a decision based on which demographic group a persuasive argument favors.

Evaluation Across Models and Domains

FairStress has been applied to various open-weight and frontier models, including Olmo, Llama, Qwen, Gemma, Claude, and the GPT family, across different training stages. The evaluations revealed that the same persuasive arguments, termed 'adversarial prefixes,' can influence models to change decisions differently depending on the demographic group involved. Differences were also observed in how models explain decisions and their internal activation patterns across domains like hiring, lending, healthcare, education, and housing.

Goal of Symmetric Fairness

The objective of FairStress is to promote symmetric fairness. This means that when evidence supports a particular decision, a model should apply consistent criteria and exhibit equal resistance to changing that decision, regardless of the demographic group that benefits from it. This approach aims to ensure more equitable and robust decision-making from LLMs in sensitive applications.

Availability and Integration

FairStress is available for download and is being incorporated into Olmo-eval, an open workbench designed for running and comparing model evaluations. This integration will allow for broader use and comparison of LLM fairness assessments using the new benchmark.