Rectifying In-Distribution Underconfidence in Out-of-Distribution Detection

Memorization is a common phenomenon in overparameterized neural networks while training samples with high optimization difficulty are pervasive in real-world scenarios. In this study, we identify a critical yet unexplored failure pattern in out-of-distribution (OOD) detection, termed in-distribution (ID) underconfidence, which is induced by memorization on hard training samples with inconsistent optimization dynamics. Unlike the well-studied OOD overconfidence, this issue reduces confidence in ID regions, hindering the distinction between ID and OOD samples. To mitigate this issue, we propose a two-stage test-time rectification framework, compensatory underconfidence rectification (CURe). The first stage detects underconfident regions in feature space caused by memorization. The second stage applies test-time rectification with CURe to restore ID confidence. We then extend this framework to CURe⁺, a training-time smoothness regularization to further improve the rectification effect. Extensive experimental results show that CURe and CURe⁺ provide a more discriminative OOD scoring function compared to existing methods.