Detecting out-of-distribution (OOD) data has become a critical component in\nensuring the safe deployment of machine learning models in the real world.\nExisting OOD detection approaches primarily rely on the output or feature space\nfor deriving OOD scores, while largely overlooking information from the\ngradient space. In this paper, we present GradNorm, a simple and effective\napproach for detecting OOD inputs by utilizing information extracted from the\ngradient space. GradNorm directly employs the vector norm of gradients,\nbackpropagated from the KL divergence between the softmax output and a uniform\nprobability distribution. Our key idea is that the magnitude of gradients is\nhigher for in-distribution (ID) data than that for OOD data, making it\ninformative for OOD detection. GradNorm demonstrates superior performance,\nreducing the average FPR95 by up to 16.33% compared to the previous best\nmethod.\n