Where Does Reasoning Break? Step-Level Hallucination Detection via Hidden-State Geometry

arXiv·medium signal

Paper frames hallucination as a property of hidden-state trajectories during a single forward pass. Correct reasoning follows a stable manifold of locally coherent transitions; the first error appears as a localized excursion measurable via transport geometry. Unlike existing trace-level detectors that give one confidence score per output, this method localizes the exact step where reasoning breaks without requiring multiple sampled completions.