ODD: Optical Depth Dynamics as a Universal Framework for Neural Network Diagnosis Across Vision and Language

Optical Depth Dynamics (ODD) is a universal diagnostic framework for analyzing internal failure modes of deep neural networks across both vision and language modalities. While modern neural networks achieve high benchmark performance, they often fail silently under distributional shifts—such as geometric transformations in vision or hallucination during language generation—without any reliable warning from output-level metrics. ODD addresses this limitation by treating neural networks as hierarchical dynamical systems and introducing depth as a diagnostic dimension. For vision models, depth corresponds to spatial network layers (from early “UV” feature extraction to late “IR” decision layers). For language models, depth corresponds to temporal progression across generation steps. By measuring representational drift relative to a calibrated in-distribution baseline, ODD reveals where, when, and how failures emerge inside the model. The framework functions analogously to medical imaging: just as MRI scans reveal latent pathology layer by layer, ODD scans neural networks across spatial or temporal depth to expose hidden instabilities before output behavior degrades. This enables internal failure localization, classification, and early warning—capabilities absent from conventional robustness and OOD detection methods. Using ODD, this work identifies several recurring failure signatures: Mid-Layer Bulge — characteristic of spurious correlations and shortcut learning, where instability peaks in intermediate representations. Global Collapse — observed in Vision Transformers under geometric stress, caused by brittle positional encodings and domain mismatch. Avalanche Effect — gradual error accumulation in convolutional architectures, leading to graceful but measurable degradation. Temporal Decoupling — a language-model–specific signature in which representational drift increases across generation steps, strongly correlated with hallucination onset. The paper further demonstrates that architectural interventions can heal internal representations. In particular, Holo-Pruning—a structured pruning strategy—reduces spatial drift by up to 54% in ResNet models under rotation stress, showing that pruning improves internal stability rather than merely reducing parameter count. ODD is lightweight (