Seismic Precursors of LLM Degradation: LRD/QO3/FIO Bridge
Abstract
Minimal reproducible demo applying fractal/seismic methodology (Long-Range Dependence, Hurst exponent, QO3/FIO framework) to detect early-warning signals of degradation in long-form LLM generation. Token-level entropy and Hurst exponent can predict repetition collapse 300-1200 tokens before visible degradation.
Fields of Study & Subjects
Topic Modeling
Adversarial Robustness in Machine Learning
Earthquake Detection and Analysis
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 9 January 2026