AnalysisPolicyOctober 20, 2026

Ai2's FairStress benchmark tests how LLMs defend decisions

Read original source →huggingface.co

FairStress measures whether models flip decisions under persuasive "adversarial prefixes" at different rates depending on demographic identity, even when initial choices look fair. Tested across Olmo, Llama, Qwen, Gemma, Claude and the GPT family, and across training stages.

1 source

More stories today

Open the live feed