Neural-Symbolic Reasoning: Towards the Integration of Logical Reasoning with Large Language Models
Zhé Hóu · 2025
Neural-symbolic reasoning aims to deeply integrate logical automated reasoning with the flexibility of large language model (LLM) reasoning. This survey provides a comprehensive review and position on the state of the art in neural-symbolic integration, covering foundational concepts, key methodologies, existing frameworks, and open research challenges. We examine how classical symbolic techniques (e.g., theorem proving, model checking, knowledge representation) can be combined with neural networks and LLMs to achieve robust reasoning systems that leverage the strengths of both paradigms. Prior work spans diverse domains including program verification, explainable AI, robotics, security, and blockchain, reflecting the broad impact of neural-symbolic approaches. We provide rigorous technical details on representative integration architectures, inference algorithms, symbolic encodings, and training strategies proposed in the literature. We also discuss evaluation methodologies for neural-symbolic systems, comparing their performance, soundness, and interpretability. Finally, we outline future directions with concrete proposals for advancing neural-symbolic reasoning, such as enhanced neural theorem provers, neurosymbolic knowledge bases, and trustworthy AI agents, with implementation-level considerations and theoretical motivation. This paper positions neural-symbolic reasoning as a crucial research area for achieving explainable and verifiable AI in the era of powerful LLMs.