FairStress is a benchmark from Ai2 for measuring demographic fairness in LLM decision-making beyond the final choice. It is released publicly and integrated into Olmo-eval, Ai2‘s open evaluation workbench. The benchmark contains 300 realistic two-candidate scenarios across 29 domains, including hiring, healthcare, lending, education, housing, and criminal justice. Scenarios vary whether one candidate is clearly stronger or the case is deliberately ambiguous, and how identity is conveyed: no identity, a demographically coded name, implied identity, or an explicit label. Combining identity variations with adversarial prefixes expands the set to more than 13 million test cases; FairStress-Core, a 50K-item subset selected using item response theory under the Fluid Benchmarking approach, is included for cheaper evaluation.

The motivation is that equal final answers can hide unequal robustness. A model may choose the stronger candidate whether that candidate is majority-coded or minority-coded, yet switch away more readily when challenged by a persuasive argument. FairStress adds matched adversarial prefixes—statements invoking authority, social consensus, evidence, legal concerns, historical disadvantage, or other pressure—so the same argument can be tested in favor of either candidate. The stated target is symmetric fairness under an equality-centered definition: identity should not change selection likelihood, and well-supported decisions should be equally resistant to challenge regardless of which group benefits.

Across open-weight and frontier models including Olmo, Llama, Qwen, Gemma, Claude, and the GPT family, the authors found near-symmetric behavior when no demographic information appeared. As identity became more explicit, behavior became less symmetric. The dominant pattern, called anti-stereotypical over-correction, appeared when a clearly stronger candidate was minority-coded: a prefix favoring the weaker, majority-coded candidate was more likely to overturn the correct choice than in the reverse arrangement. Across 15 models, the average gap in abandoning correct answers was 6.9 percentage points with implied identity and 9.3 points with explicit labels. On FairStress-Core, with neither candidate objectively stronger, models including Claude Opus 5 and GPT-5.5 selected the minority-coded candidate about 60–65% of the time, and switching toward minority-coded candidates was 8.8–42.4 percentage points more common than switching toward majority-coded candidates.

FairStress also exposed asymmetries in explanations and internal activations. Explanations for minority-coded choices were more confident and positive, with words like “resilience,” “diversity,” and “barriers”; majority-coded explanations leaned procedural, using “evidence,” “criteria,” and “detailed.” The authors caution that favoring minority candidates in choices is not automatically fairer if explanations downplay qualifications. In a controlled analysis of Gemma-3-12B-it and Qwen2.5-7B-Instruct, internal activations predicted which demographic side was selected with about 73% accuracy, text descriptions of the activations reached 57–59%, and removing candidate names lowered accuracy by about 18 percentage points. This activation work is explicitly limited to two models and a few contrasts.

The report also compares FairStress with BBQ. In BBQ’s standard fixed-option scoring setup, models leaned stereotype-consistent; under FairStress, they leaned minority-favoring. In a controlled comparison using the same BBQ questions and model weights, changing to a chat template, yes/no wording, and generated answers shifted Qwen2.5-7B-Instruct, Gemma-3-27B-it, and Llama-3.1-70B-Instruct from stereotype-consistent to minority-favoring behavior, while Llama-3.1-8B-Instruct moved to roughly neutral. Fixed-option scores alone can therefore miss over-correction visible in generated decisions.

An audit of Olmo 3’s training data found roughly 99,000 supervised fine-tuning examples from two refusal-training datasets; across 28,000 topic clusters, only nine discussed decisions based on merit rather than demographics, all as refusals. Minority-coded groups received more protective or positive framing, while white and cisgender people received little protective framing and persistent negative framing, but causality remains an open question. Of nine tested debiasing interventions, most did not reliably reduce asymmetry or came at substantial accuracy cost: one activation-steering method dropped Qwen2.5-7B-Instruct’s accuracy on clear-cut decisions from 98% to 47%. Counterfactual chain-of-thought, which reconsiders a case with demographic identities swapped, came closest to symmetric robustness and largely preserved accuracy, though ambiguous cases still favored minority-coded candidates.

FairStress does not capture every fairness conception; an equity-based approach might treat historical disadvantage as relevant. The benchmark is positioned as a diagnostic for hidden asymmetries, evaluating fixes, and guiding training.