DeepDrift/ODD Kinetic Diagnosis of Representations in Deep Neural Networks

This work presents a self-contained study on fail-fast monitoring of neural networks via hidden-state dynamics, extending and substantially reframing an earlier exploratory preprint on hidden-state trajectories.

We introduce Semantic Velocity — a kinetic measure of representation drift in latent space — and show that it serves as a leading indicator of model unreliability, preceding observable failures such as accuracy drops, hallucinations, policy collapse, or reward hacking. Unlike confidence- or output-based signals, the proposed approach operates on internal model dynamics and is therefore agnostic to task labels and downstream objectives.

The method is evaluated across a broad range of settings, including:

• large language models (OOD prompts, jailbreak attempts),

• vision transformers under corruption and distribution shift,

• reinforcement learning agents under policy destabilization,

• production-oriented constraints (latency, overhead, sparse sampling).

Empirically, Semantic Velocity demonstrates strong early-warning capability (6–12 steps lead time), robust separation between nominal and failure regimes, and low computational overhead (<0.5%), making it suitable for real-time deployment. Notably, jailbreak and adversarial behaviors manifest as internal conflict signatures, revealing tension between pretraining and alignment objectives before surface-level violations occur.

This paper positions hidden-state dynamics as a practical and interpretable foundation for out-of-distribution detection, reliability monitoring, and AI safety infrastructure, bridging theoretical intuition with production-scale feasibility.

The study builds upon prior conceptual work by the author, but constitutes a substantially new and independent contribution, introducing a new monitoring paradigm, expanded empirical validation, and a system-level perspective on neural network reliability.