Modeling Decision Thresholds for Out-of-Distribution Detection

Abstract

Out-of-Distribution (OOD) detection is essential for maintaining the reliability of machine learning systems when deployed in dynamic real-world environments where input distributions may shift over time. This study evaluates multiple threshold modeling approaches, including confidence-based scoring, latent-space distance evaluation, energy-based scoring, and adaptive threshold recalibration. Experimental results demonstrate that confidence-based thresholds are insufficient for distinguishing unfamiliar samples due to poor uncertainty calibration. Distance-based and energy-based scoring models provide more robust separation between in-distribution and OOD inputs by leveraging the geometric structure of learned feature manifolds. Furthermore, adaptive thresholding strategies maintain stable detection performance under distributional drift, outperforming fixed thresholds in evolving operational contexts. These findings highlight the importance of geometry-aware and dynamically tunable threshold models for reliable deployment of neural systems in production settings.