Convergence of LLM Integration and Edge Intelligence in Decentralized IoT Ecosystems

Rahul Rishi Sharma
Khushboo Bhatia
Nagesh Gulkotwar
Amanpreet Kaur
Shreya Gupta
2026
Google Scholar

Abstract

This paper details the fundamental transition from conventional Internet of Things (IoT) frameworks to a decentralized Internet of Agents (IoA), where Large Language Models (LLMs) serve as the central reasoning kernels for Edge Intelligence and Physical AI. This enables autonomous agents to perceive heterogeneous environments, manage complex goals over extended temporal horizons, and collaborate across distributed network topologies, thus addressing the versatility-efficiency-portability trilemma. This paper explores core architectural challenges, including overcoming the modality gap between LLM-trained textual corpora and numerical sensor telemetry through techniques like discrete tokenization, multisensory alignment, and visualizing time-series data. To ensure long-horizon execution, the paper proposes a novel hierarchical planning framework that decomposes high-level objectives into executable subgoals, complemented by Continuum Memory Architectures (CMA) for persistent identity and context. Furthermore, it outlines crucial optimization techniques, such as quantization and structured pruning for resource-constrained edge deployment, and stresses the need for hardening security protocols like MQTT with authentication and fine-grained access control lists. The integration of these sophisticated agent architectures with IoT moves the industry past reactive automation, setting the foundation for an autonomous behavior where strategic focus should lie on refining tokenization, adopting secured protocols, and utilizing domain-grounded benchmarks like AssetOpsBench.

Research Areas

×