A Roadmap Toward Neurosymbolic Approaches in AI Design

Neurosymbolic Artificial Intelligence (NS-AI) integrates two central paradigms in artificial intelligence (AI). The first is connectionist (or sub-symbolic) AI, which enables scalable and efficient statistical learning. The second is symbolic reasoning, which provides formal reasoning and explainability. This study conducted a systematic review of NS-AI architectures published between 2013 and 2024, synthesizing the structural patterns and integration strategies in 319 peer-reviewed publications. From this corpus, 18 representative approaches were selected based on their methodological rigor, symbolic fidelity, and conceptual novelty. These approaches were subsequently classified into three architectural paradigms: Sequential, Multi-Integration, and Hybrid. Based on this classification, a five-stage symbolic integration framework is proposed, comprising (i) data preprocessing, (ii) neural–symbolic embedding, (iii) incorporation of domain knowledge, (iv) logical reasoning modules, and (v) symbolic postprocessing. The findings indicate that multistage symbolic integration, particularly when embedded within neural layers and reasoning components, improves system explainability, robustness, and semantic alignment. Comparative analysis further reveals a fundamental trade-off: architectures grounded in formal logic exhibit higher precision and verifiability, whereas those employing representational-symbolic languages yield greater interpretability and flexibility. The resulting generic architecture provides a modular design blueprint for future NS-AI systems, supporting scalable, transparent, and semantically grounded AI development.