Digital Brain: A Modular Cognitive Architecture as an Alternative to the Large Language Model Paradigm

The current Large Language Model (LLM) paradigm based on the Transformer architecture adopts a "monolithic chemical plant" approach, compressing perception, memory, reasoning, decision-making, and generation all into a single set of parameters. While achieving remarkable general-purpose capabilities, this paradigm suffers from fundamental deficiencies: hallucinations, unreliable reasoning, extreme energy consumption, lack of auditability, and inability to incrementally update. This paper proposes the Digital Brain — a modular cognitive architecture consisting of three independent modules: Cognition, Logic, and Knowledge Base. Rather than relying on a massive neural network as a universal reasoning engine, the Digital Brain achieves systematic substitution of the LLM paradigm through standardized interface collaboration. Key design features include: (1) the Cognition module handles input understanding and domain routing, requiring only small models or rule engines; (2) the Logic module performs deterministic reasoning, planning, and decision-making based on symbolic systems, completely eliminating hallucinations; (3) the Knowledge Base module provides explicit fact storage and experience accumulation, supporting lifelong learning. The three modules are connected through a feedback loop, enabling continuous evolution during operation. Comparative analysis shows that compared to equivalent-capability LLMs, the Digital Brain reduces parameter count by approximately 3-4 orders of magnitude (~50M vs 200B), reduces power consumption by approximately two orders of magnitude (~2W vs 700W), can be deployed on a single embedded board, and provides fully auditable and traceable reasoning processes.Author contact: Chen Bo, 1392324000@qq.com