What if we train a classifier on one data environment and apply it on another environment? In #ICML2021 , we show that its prediction results are informative of the unstable correlations between the input and the output. This allows us to effectively train a stable classifier.