Training-free Out-of-Distribution (OOD) detection that beats state-of-the-art by aggregating features across intermediate network layers.
March 26, 2026
Original Paper
Prototype Fusion: A Training-Free Multi-Layer Approach to OOD Detection
arXiv · 2603.23677
The Takeaway
It challenges the conventional wisdom that the penultimate layer holds the most discriminative information for OOD tasks. By fusing 'prototypes' from multiple layers, the method improves AUROC by up to 4.4% and reduces False Positive Rates by 13.5% without any model retraining.
From the abstract
Deep learning models are increasingly deployed in safety-critical applications, where reliable out-of-distribution (OOD) detection is essential to ensure robustness. Existing methods predominantly rely on the penultimate-layer activations of neural networks, assuming they encapsulate the most informative in-distribution (ID) representations. In this work, we revisit this assumption to show that intermediate layers encode equally rich and discriminative information for OOD detection. Based on thi