Spectral Analysis of Training-Free LLM Text Detection

This study analyzes spectral energy in log-probability trajectories to detect machine-generated text. It identifies 'generative vitality'—fluctuations in token choice—as a key differentiator between human and model-generated content, particularly in long-form text.