Large Language Models for Future Internet Ecosystems: A Taxonomy-Based Review of Smart IoT, Edge Intelligence, and Autonomous AI

Large Language Models (LLMs) are emerging as key enablers of Future Internet ecosystems, supporting intelligent, adaptive, and context-aware services across distributed cyber-physical environments. Beyond traditional natural language processing, LLMs are increasingly integrated into Smart Internet of Things (SIoT) systems to enable semantic interoperability, edge intelligence, multimodal interaction, and autonomous service orchestration. This paper presents a taxonomy-based review of LLM architectures, training paradigms, deployment strategies, and emerging applications within Future Internet infrastructures. The review classifies existing studies by deployment environment, architecture, training strategy, accessibility, and application scope, and analyses the role of LLMs in intelligent IoT environments, with emphasis on edge-based reasoning, agentic AI, human-centric automation, and context-aware decision-making. Key challenges are examined, including scalability, inference latency, privacy, trustworthiness, security, hallucination, and energy efficiency in resource-constrained environments. A comparative analysis of representative LLMs is presented, based on deployment feasibility, multimodal capability, accessibility, and suitability for distributed intelligent services. The originality of the review lies in conceptualizing LLMs as cognitive middleware that provides semantic, reasoning, and coordination capabilities across Smart IoT infrastructures. Finally, future research directions are highlighted, including decentralized AI architectures, digital twins, the Model Context Protocol, retrieval-augmented generation, multimodal sensing, and autonomous agent-based ecosystems.