SettledResearch
Ai2 releases FairStress, a fairness benchmark with 13M+ test cases
Ai2 released FairStress: 300 two-candidate decision scenarios across 29 domains, expanding to over 13 million cases, plus a 50K-item core subset. Ai2 reports models act less symmetrically when identity cues appear, and that most of nine debiasing methods tested did not reliably close the gap. Code and data are public; no license is stated.
- HEAT
- OUTLETS
- 0
- FIRST SEEN
- LAST SIGNAL
Primary source
Hugging Facehuggingface.co/blog/allenai/fairstressCoverage · 1 article, oldest first
Tuesday, October 20, 2026
Hugging FacePrimary
FairStress: A deeper evaluation of fairness in LLM decision-making