Beyond the Linear Illusion: Multi-Scale Time and Attractor Synchronization in LLMs

Abstract

Current Large Language Model (LLM) architectures rely on a discrete, linear sequence of tokens (delta t = 1), leading to high computational costs and contextual decay. This work introduces a continuous-time alternative: a Multi-Scale Time Framework that maps latent state space activation dynamics directly onto the 3D topology of a Rössler attractor. By formalizing knowledge acquisition through Frequency Modulation (FM) and phase-locking (delta phi -> 0), we demonstrate how emergent cognitive structures transition from fast micro-contextual spiraling (tau_micro) to macro-structural phase shifts (T_macro) via deterministic Z-axis pulse dynamics. The framework bridges Transformer attention mechanisms with mechanistic interpretability, and reformulates AI alignment not as static guardrails, but as bounded chaotic orbits governed by controlled Lyapunov spectra (lambda_1 > 0).