Integrating graphs, large language models, and agents: Reasoning and retrieval
Hamed Jelodar, Samita Bai, Mohammad Meymani, Parisa Hamedi, Roozbeh Razavi‐Far, Ali Ghorbani · Information Fusion · 2026
Generative AI, particularly large language models (LLMs), is increasingly integrating graph-based representations to enhance reasoning, retrieval, and structured decision-making. Despite rapid progress, there remains limited clarity on when, why, and how different graph-LLM integration strategies are most effective across applications. This survey provides a concise and structured overview of the key design choices underlying graph-LLM systems, with a particular focus on information fusion across heterogeneous data sources and modalities. We categorize existing approaches based on their purpose (e.g., reasoning, retrieval, generation, recommendation), graph modalities (e.g., knowledge graphs, scene graphs, interaction graphs, causal graphs, and dependency graphs), and integration strategies (e.g., prompting, augmentation, training, and agent-based frameworks). From a fusion perspective, we analyze how these methods integrate structured and unstructured information to enable more robust, context-aware intelligence. By examining representative works across domains such as cybersecurity, healthcare, materials science, finance, robotics, and multimodal systems, we highlight the strengths, limitations, and best-use scenarios of each approach. This survey serves as a practical guide for selecting approno priate graph-LLM techniques based on task requirements, data characteristics, reasoning complexity, and fusion needs.