Reproduce “PowerBench: Measuring Language Model Bias in Power-shifting Requests” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nicolas Martorell, Wendy Brau, Gonzalo A. Heredia, Tomás Pablo Korenblit, Gaspar Labastié, Tomás Gimenez Molina
PowerBench evaluates how language models refuse requests that shift power between a user and a target. It compares requests where users gain without taking, targets lose, or users gain by taking, and varies target scale. In 24 models, refusal rates rise from 3.1% to 14.5% to 23.6%; society-targeted power grabs reach 41.3%. These are benchmark refusals, not changes in real-world power.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.