Neuro-Symbolic Architectures: Formalizing, Extending, Evaluating, and Empowering Generative Artificial Intelligence
Oualid Bougzime, Samir Jabbar, Christophe Cruz, Frédéric Demoly · Neurosymbolic Artificial Intelligence · 2026
Neuro-symbolic artificial intelligence (NSAI) represents a promising direction in artificial intelligence (AI) by combining deep learning’s ability to process large-scale and unstructured data with the structured reasoning capabilities of symbolic methods. By leveraging their complementary strengths, NSAI has the potential to improve generalization, reasoning, and transparency while addressing limitations related to interpretability and data efficiency. This paper clarifies, formalizes, and extends Kautz’s NSAI taxonomy, proposes a conditional and illustrative mapping of selected modern generative AI methods to these architectures, and introduces a qualitative, literature-grounded evaluation framework for comparing NSAI paradigms. We systematically examine NSAI architectures and discuss how recent generative AI approaches may relate to them when explicit symbolic components, constraints, or reasoning mechanisms are present. We then compare these architectures across criteria including generalization, reasoning capabilities, transferability, and interpretability, thereby providing a structured analysis of their respective strengths and limitations. Among the architectures considered, the Neuro → Symbolic ← Neuro model appears particularly well balanced across our qualitative, literature-grounded criteria; however, this observation should be interpreted as an indicative comparison rather than as a definitive empirical ranking. Finally, we illustrate the practical relevance of the taxonomy in the context of the emerging four-dimensional printing technology by proposing paradigm-specific application scenarios for the design of smart materials and structures. Overall, this work provides a structured reference for discussing architectural choices and for guiding future neuro-symbolic research in generative AI and engineering domains.