From External Grounding to Collective Intelligence: A Comprehensive Survey on Factuality, Knowledge Graphs, and Multi-Agent Orchestration in Large Language Models

Abstract: Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) demonstrate remarkable generative capabilities but suffer from persistent hallucinations, factual drift, and reasoning degradation in high- consequence domains. While canonical Retrieval-Augmented Generation (RAG) mitigates parameter staleness, it introduces critical vulnerabilities including neural ranking miscalibration, numerical insensitivity, and channel noise cascading. This survey provides an exhaustive, taxonomy-driven synthesis of recent advances across four synergistic research pillars: evidential retrieval calibration, information-theoretic entropy bounds, structured knowledge graph (KG) grounding, and multi-agent collaborative orchestration with asymmetric verification against hallucination snowballing. Adhering strictly to PRISMA 2020 guidelines across twenty primary benchmarked architectures and over eighty foundational sources, we formalize the mathematical mechanics of epistemic uncertainty and cascading channel capacity. We present multidimensional comparative matrices evaluating accuracy, computational complexity, and inference latency, systematically analyze open research frontiers, and establish concrete engineering design trade-offs for deploying reliable, graph-grounded multi-agent intelligence in mission-critical environments.

Keywords: Large Language Models, Retrieval-Augmented Generation, Knowledge Graphs, Hallucination Mitigation, Evidential Deep Learning, Multi-Agent Systems, Information Theory