Beyond Hallucinations: Toward Structural Guidance in Large Language Models

This paper is part of a broader research program on the structural analysis of complex and living systems, as outlined in the “Document Zero”. Large Language Models (LLMs) have demonstrated remarkable capabilities, yet they exhibit persistent limitations such as hallucinations, instability, and lack of global coherence. These issues are typically attributed to data limitations or alignment challenges, but such explanations remain incomplete. This work proposes a structural interpretation of these limitations, arguing that current LLMs function primarily as high-capacity propagation systems lacking an explicit internal layer for regulation and organization. As a result, coherence is not maintained as a global invariant, leading to generative drift and instability. To address this gap, the paper introduces the concept of a structural guidance layer, designed to operate as an intermediate level between generative processes and coherent cognitive organization. This layer relies on constraint-based regulation, operator-based transformations, and the identification of structural signatures to guide system behavior. The proposed framework does not replace existing architectures but complements them by enabling guided generation, improved coherence, and adaptive reconfiguration. This contribution constitutes an application of a broader systemic framework to artificial intelligence and is conceptually connected to ongoing work on distributed cognitive architectures (DCNM, currently under peer review), which aims to formalize the structural principles underlying constraint-based cognition.