Theoretical Bounds on LLM Token Generation via Dynamical Systems

This research formalizes LLM response classification as a binary hypothesis test between stochastic linear dynamical systems. It establishes a fundamental accuracy floor for classifiers that ignore token dynamics and proves that misclassification probability decays exponentially with sequence length.