Architectural Taxonomy of LLM and Knowledge Graph Hybrids for Trustworthy AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Large language models (LLMs) and knowledge graphs (KGs) are increasingly combined to improve factual grounding, reasoning, explainability, and trustworthiness. Existing studies primarily treat such systems as instances of LLM–KG integration, paying little attention to the hybrid architecture itself. Consequently, ethical, privacy, and bias (EPB) analyses rarely consider how different hybrid architectures influence the emergence, propagation, amplification, or mitigation of risks. In this paper, we argue that hybridization should be treated as a first-class object of study. We provide an initial algebraic formalization of hybrid systems and introduce a general three-dimensional taxonomy of computational hybrids that distinguishes (i) construction mode (conventional or self-generating), (ii) interaction topology (sequential, interactive, mixed, managed, or cross-validating), and (iii) blending realization (static or adaptive learnable blending). Together these dimensions define a structured design space of twenty representative LLM–KG hybrid architectures. Building on this foundation, we derive a hybrid-aware taxonomy of EPB risks by systematically analyzing the emergent capabilities, as well as the emergent and mitigated EPB risks, associated with each architecture. Our analysis demonstrates that the trustworthiness of an LLM–KG system depends not only on its constituent components, but also on how they are hybridized. We further introduce cross-validating hybrids, in which LLMs and KGs continuously verify one another, and learnable blending models that transform hybridization into an optimization problem, enabling the automatic synthesis of trustworthy hybrid architectures. Although motivated by LLM–KG systems, the proposed framework is domain-independent and provides a conceptual foundation for designing, comparing, optimizing, and evaluating heterogeneous AI systems.