Runtime Fisher Spectral Sensitivity for Early Hallucination Detection

Large language models (LLMs) sometimes "hallucinate" — they confidently produce statements that are factually wrong or made up. This paper from researchers at the University of Amsterdam introduces a new method to detect hallucinations while the model is still generating text (i.e., at "runtime"), rather than after the fact. The key innovation is using Fisher's method (a statistical technique for combining p-values) on the model's internal spectral (frequency-domain) activations — signals from the model's internal layers analyzed in a frequency representation — to flag early signs of hallucination. Fisher's method is normally used in meta-analysis to combine results from multiple independent tests; here it is applied to combine "evidence scores" across different layers of the neural network. The approach is "spectral" because it transforms hidden-state signals into the frequency domain (like analyzing sound waves) to detect patterns the raw numbers might miss. This is part of a growing push to make LLMs more reliable by catching their mistakes as they happen, without needing external databases or multiple model runs.