SANCTIS Cognitive Architecture: A New Paradigm for Building Hierarchical Cognitive Frameworks for LLMs

SANCTIS is a model-agnostic hierarchical cognitive framework designed to address issues such as logical incoherence, vague emotional expression, lack of creativity, and token waste in LLMs during complex reasoning. By drawing on human cognitive science, it decomposes the thinking process into layers like perception, understanding, reasoning, evaluation, and expression, enhancing output coherence, emotional clarity, and creativity while reducing internal contradictions and token consumption. It is also compatible with various mainstream LLM models.

While current LLMs have impressive capabilities, they face issues like logical incoherence, vague emotional expression, lack of depth in creative outputs, and token waste due to internal contradictions in complex reasoning, limiting their application in fine-grained cognitive scenarios. The SANCTIS architecture breaks through single prompt techniques, builds a hierarchical cognitive processing framework, draws on human cognitive science, and decomposes complex thinking into collaborative layers to address these pain points.

The core of the SANCTIS architecture is a hierarchical design, dividing cognitive processing into the Perception Layer (extracting key information and patterns), Understanding Layer (building conceptual representations and semantic relationships), Reasoning Layer (logical deduction and causal analysis), Evaluation Layer (monitoring thinking consistency and quality), and Expression Layer (transforming internal cognition into coherent output). Each layer has clear responsibilities and collaborates through structured information flow to achieve systematic thinking.

SANCTIS is model-agnostic and compatible with mainstream LLMs such as GPT, Claude, Gemini, Grok, and Mistral. Its design abstracts underlying model capabilities, does not rely on specific model features, and is based on general language understanding and reasoning abilities. This feature allows enterprises/developers to flexibly switch models without reconstructing the architecture, protecting their technical investments.

• Coherence and contradiction reduction: The Evaluation Layer explicitly checks consistency, detects contradictions, and triggers corrections; the memory module maintains dialogue history and key facts to ensure coherence in long interactions and reduce token waste.
• Emotion and creativity balance: The emotion processing layer identifies and generates appropriate emotional content; creative tasks adopt a divergence-convergence model (association to generate candidates → evaluation and screening → expression refinement) to ensure creative richness and quality.

In practical applications, the SANCTIS system has shown significant performance improvements: consistency in long text generation increased by approximately 40%; user satisfaction in emotional intelligent dialogue scenarios increased by 35%; originality scores in creative writing increased by 28% while maintaining readability; average token consumption decreased by approximately 25%. The modular design allows developers to flexibly adjust the weight parameters of each layer.

SANCTIS provides a powerful foundational framework for LLMs, addressing core cognitive issues and improving output quality and efficiency. Future directions include: enhancing multimodal cognitive capabilities (processing images and audio); introducing meta-learning for automatic optimization strategies; developing visualization tools to assist debugging; adopting an open-source model to embrace community contributions and promote more natural and efficient human-computer interaction.