Participants had to predict the questions someone could solve, given the information that this person solved (or not) another question.
Their predictions were highly accurate compared to our benchmark data and matched those of an optimal Bayesian model. Even participants bad at math were accurate!
