Hugging Face AI Open Source Blog: FairStress: A deeper evaluation of fairness in LLM decision-making

AI2 has released the FairStress benchmark, expanding 300 scenarios into over 13 million test cases to measure how large language models handle fairness when facing adversarial prefixes in high-stakes decisions. The evaluation found that models are more likely to abandon their correct decision to support minority-group-coded candidates when identity information is explicitly provided, with an average abandonment gap of 6.9 percentage points (implicit identity) to 9.3 percentage points (explicit labels). The benchmark is now available for download and has been integrated into Olmo-eval. Among the nine debiasing methods tested, only counterfactual chain-of-thought reasoning managed to narrow the gap while maintaining accuracy.