GulliBench: Measuring Skepticism in Frontier Models

I liked approach to evaluating AI behavior, the fact that additional reasoning doesn’t improve performance is quite interesting!

interesting approach