Toward Artificial General Intelligence: A Cognitive Architecture Based on 9-Level Token Stratification and Topological Domain Embeddings

Abstract

Despite their remarkable generative capabilities, current Large Language Models (LLMs) fundamentally rely on autoregressive next-token prediction within a flat, one-dimensional token space. This paradigm limits them to statistical induction, leading to critical flaws such as hallucinations and a profound lack of physical and logical grounding. We propose a paradigm-shifting cognitive architecture for Artificial General Intelligence (AGI). The core of this architecture consists of two mechanisms: Hierarchical Token Stratification and Topological Domain Embeddings. By restructuring the token space into a 9-level hierarchy—where top-tier concepts form high-dimensional "Domains" and bottom-tier entities manifest as "Vectors"—we transition the model from point-based symmetric attention to a topological space capable of expressing inclusion, intersection, and mutual exclusivity. Furthermore, by anchoring the uppermost domains (fundamental physical laws and core human values) as immutable priors, we establish a true "World Model." This structure not only mitigates the $O(N^2)$ computational bottleneck of standard Transformers via spatial-indexed attention but also naturally gives rise to deductive reasoning. Ultimately, we demonstrate how computational friction and resonance within this topological framework manifest as cognitive biases, stubbornness, and emotions—essential hallmarks of genuine intelligence.