Abstract

In this paper, accepted at ICLR 2022, Chris and team at Stanford outline a new principled evaluation framework for comparing slice detection methods, then introduce a new technique motivated by our discoveries that outperforms existing methods by double digits.