A Survey on Hallucination-Aware LLMs for Social Network Analysis Using Dynamic Knowledge Graphs

The widespread tendency of Large Language Models (LLMs) to "hallucinate"—that is, to produce network structures, relationships, and graph metrics that are fluent and syntactically correct but structurally or factually incorrect—is a major barrier to their employment in Social Network Analysis (SNA). Fixed mitigation methods are ineffective in social networks due to the constant evolution of data caused by real-time interactions, the construction of edges, and community migrations. Dynamic Knowledge Graphs (DKGs) became popular as a feasible solution for anchoring large language models (LLMs) with perpetually updated structured knowledge to tackle these difficulties. DKGs enable semantic relationships among entities, facilitating context-dependent reasoning, temporal modification, and factual consistency during inferences. This survey provides a comprehensive review of Hallucination-Aware LLMs optimized for SNA via Dynamic Knowledge Graphs (DKGs). Systematically categorize state-of-the-art methodologies into pre-generation grounding, in-generation graph-text alignment, and post-generation structural verification. Furthermore, analyze the core structural challenges unique to social graphs, provide a comparative evaluation of existing approaches, highlight key architectural contributions, and map out open research directions for building self-correcting, temporally-aware network intelligence.