New training-free method uses multiple neural network layers to better detect out-of-distribution data in AI systems

Researchers challenge the assumption that only the penultimate layer of neural networks is useful for OOD detection

The approach aggregates features from multiple convolutional blocks to create class-wise prototypes without requiring additional training

Uses cosine similarity scoring between test features and learned prototypes to identify out-of-distribution samples

The model-agnostic method applies L2 normalization to form compact prototypes that capture class semantics across layers

Designed for safety-critical applications where reliable OOD detection is essential for AI robustness