Distributed Computing with Language Models at the Edge: A Framework and Prototype
Dinesh Kumar Karthikeyan, Roopesh Kumar Shanmugasundaram, Anna-Sofia Paavonen, Tommi Mikkonen, Niko Mäkitalo, Ronny Seiger · 2025
The evolution of cyber-physical systems (CPS) necessitates new methods to manage growing complexity and dynamic interactions between physical and digital domains. This paper presents the Edge Multimodal Intelligence Network on Devices (EdgeMIND) – a cognitive edge computing framework to support CPS development. EdgeMIND deploys I/O nodes powered by large language models (LLMs) across heterogeneous edge resources to process multimodal data streams. These nodes perform runtime decision-making, enabling context-aware interactions with minimal cloud reliance. The framework generates adaptive outputs by integrating retrieval-augmented generation (RAG) with Situational Awareness (SA) enhanced by topic modeling. Empirical evaluations with three LLMs on edge devices show that SA achieves up to 56% latency reduction in CPU-bound nodes and 50% in GPU-based nodes. It also reduces CPU/GPU utilization, memory usage, and thermal load. These results confirm EdgeMIND’s effectiveness for context-aware, resource-efficient multimodal processing in CPS as demonstrated in a prototype application. EdgeMIND advances the design of intelligent, efficient edge systems for responsive AI-driven services.