From the source
Ai2 introduced FairStress, a benchmark that measures demographic fairness in LLMs beyond the final decision a model makes, and is building it into its Olmo-eval workbench.
- Standard fairness benchmarks only check whether a model picks the same candidate after demographic identities are swapped, which can hide differences in how easily a model abandons that choice when challenged.
- FairStress tests this by adding adversarial prefixes, matched arguments that make the same case for either candidate, and measuring whether a model switches decisions depending on which demographic group the argument favors.
- The benchmark holds 300 two-candidate scenarios across 29 domains, including hiring, healthcare, lending, education, housing and criminal justice, with identity presented as nothing, a coded name, an implied identity or an explicit label.
- Combining those variations produces more than 13 million test cases, and a 50K-item subset called FairStress-Core, selected with item response theory, is meant to make evaluation faster and cheaper.
- Across open-weight and frontier models including Olmo, Llama, Qwen, Gemma, Claude and the GPT family, Ai2 says the same persuasive arguments can make models more or less likely to change decisions depending on which group the argument favors, with differences also appearing in explanations and internal activation patterns.
- Ai2 says its target is symmetric fairness: well-supported decisions should be equally resistant to adversarial prefixes regardless of which group benefits, a definition it notes does not cover every conception of fairness.
