DeepDrift / ODD: Hidden-State Trajectory Monitoring for Robustness in Vision, Language, and Control

Modern neural networks achieve strong performance on curated benchmarks yet fail unpredictably under distributional shift. Classical out-of-distribution detection relies on output-level confidence signals (e.g., maximum softmax probability; Hendrycks & Gimpel, ICLR 2017). However, for RLHF-tuned large language models and PPO-trained reinforcement learning agents, output confidence is often a lagging indicator: models may remain confident while internally unstable or hallucinating. This work proposes a shift from static output monitoring to internal dynamical diagnostics. We introduce Semantic Velocity, defined as the norm of consecutive hidden-state differences, as a leading indicator of representational instability that often precedes output-level failure. We present Optical Depth Dynamics (ODD), a diagnostic framework that analyzes hidden-state drift across depth—spatial depth for vision models and temporal depth for language and control. ODD provides interpretable failure signatures across modalities: • Vision: Spatial depth analysis reveals characteristic drift patterns such as Global Collapse in ViT-style architectures and Avalanche Effects in CNNs.• Language: Temporal analysis of hidden-state trajectories detects Semantic Tremor: factual generation exhibits stable, low-velocity dynamics, while hallucination is associated with high-variance drift detectable several tokens before completion.• Control: Phase-portrait analysis of reinforcement learning policies identifies Policy Panic regimes in which elevated velocity reliably precedes reward collapse (p < 0.001). All experiments follow explicit statistical protocols, including train/validation/test splits for threshold selection, bootstrap confidence intervals, and effect-size reporting. The approach is validated on Qwen-2.5-7B, TinyLlama-1.1B, LunarLander-v3, and standard vision benchmarks. The open-source library DeepDrift (pip install deepdrift) implements these diagnostics with approximately 1% computational overhead and is intended as a practical monitoring tool for robustness analysis rather than a formal physical theory.