Out-of-distribution (OOD) detection is indispensable for deploying machine learning models in the wild. Distance-based OOD detection methods are promising, but often suffer from discrepancies between the distributions learned in training vs. the distributional assumptions made in testing. This paper bridges the gap by addressing two key challenges—representation learning and OOD detection—in one coherent framework. Our proposed framework SIREN contributes two novel components: (1) a trainable lo…