Integrating Graphs, Large Language Models, and Agents: Reasoning and Retrieval

A comprehensive survey examines how Large Language Models can be effectively integrated with graph-based data structures to improve reasoning, retrieval, and decision-making across domains. The research categorizes integration approaches by purpose, graph type, and strategy, providing practitioners with guidance on selecting appropriate techniques for specific applications in healthcare, finance, robotics, and other fields.

This arXiv paper addresses a critical gap in AI research by systematizing the integration of graphs with Large Language Models, two powerful but historically separate technologies. As LLMs dominate natural language processing and graph systems excel at representing structured relationships, their combination unlocks enhanced capabilities for complex reasoning tasks. The survey's value lies in its practical taxonomy—organizing methods by purpose (reasoning, retrieval, generation, recommendation), graph modality (knowledge graphs, causal graphs, dependency graphs), and integration strategies (prompting, augmentation, training, agent-based). This structured approach reflects the field's maturation beyond experimental integration toward principled deployment strategies.

The emergence of graph-LLM hybrids responds to fundamental limitations in each technology alone. LLMs struggle with factual grounding and structured reasoning, while graphs lack semantic understanding and generative capability. By combining them, researchers address these weaknesses: graphs provide verified factual anchors and explicit relationships, while LLMs contribute language understanding and flexible reasoning. The survey's coverage across cybersecurity, healthcare, materials science, finance, and robotics demonstrates this integration's broad relevance.

For developers and researchers, this work offers immediate practical value. Understanding when to use prompting versus fine-tuning, or graph augmentation versus agent-based approaches, directly impacts project outcomes and resource allocation. The framework enables informed decision-making based on task requirements and data characteristics rather than following trends. For the broader AI ecosystem, this consolidation signals that graph-LLM integration is transitioning from novel research to established methodology, warranting standardized approaches and best practices.

• →Graph-LLM integration systematically combines structured knowledge representation with generative language understanding across four primary purposes: reasoning, retrieval, generation, and recommendation.
• →Integration strategies span prompting, augmentation, training, and agent-based approaches, each suited to different task complexity levels and resource constraints.
• →The survey maps applications across six major domains including healthcare, finance, cybersecurity, and robotics, revealing domain-specific optimization patterns.
• →Graph modality selection—whether knowledge graphs, causal graphs, or dependency graphs—significantly impacts reasoning quality and computational efficiency.
• →This survey transitions graph-LLM integration from experimental research to practical methodology with clear selection criteria for practitioners.