A Review of Large Language Models in Edge Computing: Applications, Challenges, Benefits, and Deployment Strategies

Аннотация

Large Language Models (LLMs) have achieved very good success in natural language processing, but deployment of these powerful models on edge computing devices across all domains presents unique challenges. This paper reviews the state of LLMs in edge computing, focusing on four key aspects: their emerging applications across various sectors, the technical challenges of running LLMs on resource-constrained edge devices, the potential benefits of bringing LLM capabilities closer to data sources, and effective deployment strategies to enable LLMs at the edge. We also discuss on how LLM edge deployment could offer low-latency, privacy-preserving intelligent assistance throughout a range of domains, such as healthcare, IoT, industrial automation, and more. We also look at some techniques and architectures that can overcome the limitations of edge devices, such as cloud-edge collaboration, federated learning, model compression, and on-device inference. This review identifies practical ways to integrate LLMs into edge environments by examining current practices and their trade-offs. It also provides guidance for future research to address the remaining issues in this quickly expanding field. KEYWORDS Edge computing, large language models (LLMs), Edge AI, On-Device Inference, IoT, Federated learning, Model compression, Deployment Strategies 1. INTRODUCTION Large Language Models (LLMs) like the GPT models and other transformer-based models have become a revolution in natural language processing because of their ability to interpret and generate text like humans. These models can achieve state-of-the-art performance on tasks such as language translation and summarization to answering a question and code generation. However, most LLM deployments now depend on powerful cloud servers or data center infrastructure due to the models’ powerful computational and memory requirements. On the other hand, edge computing refers to computing that is performed near the data source or end-user, for example, on local devices like smartphones, IoT sensors, or edge servers, instead of centralized cloud data centers. This paper tackles the research issue of how to efficiently bring the capabilities of large language models into edge computing environments with limited resources and possibly uneven network connectivity. Running LLMs on the network edge offers several advantages: it can enhance privacy by keeping sensitive data on- device, reduce latency by processing data locally in real-time, and allow continuous service even in areas with spotty internet. An advanced AI assistant on a smart device, for instance, could comprehend voice commands and produce https://www.academicpublishers.org/journals/index.php/ijdsml 300 AMERICAN ACADEMIC PUBLISHER immediate replies without having to reach a cloud server. Factory floor machines or autonomous robots could run language models on local controllers to understand complicated instructions on-site or make quick, split-second decisions. These examples are highly desirable for applications that require quick or reliable responses, such as healthcare diagnostics in remote clinics, autonomous vehicles navigating in real time, or military and disaster- response operations where connectivity may be limited. Enabling LLMs on edge devices is therefore important since it will help to open new intelligent possibilities in fields where timing, privacy, and dependability are top priorities. Although it is promising, achieving a very effective LLM deployment on edge devices is a big challenge. Edge devices are inherently resource-constrained, and they have much less processing power, memory, and energy capacity than cloud servers. An LLM which can run smoothly on a cloud GPU cluster might be impossible to load into a small device’s memory. Furthermore, edge environments are heterogeneous: one finds a wide variety of hardware (from smartphones and edge gateways to tiny microcontrollers) and software stacks, which makes it difficult to develop one-size-fits-all solutions. This paper aims to address these issues by reviewing and synthesizing the current research and developments on LLMs in edge computing. The goals of the paper are to first survey application domains where edge-based LLMs are being applied or could be beneficial, showing the practical use-cases driving this field; and then to discuss the technical challenges and limitations that arise when deploying LLMs on edge hardware; and finally to examine the benefits and strategies for overcoming these challenges, including methodologies like model compression, federated learning, and distributed inference. Meeting these goals helps the review to provide a thorough knowledge of how LLMs can be pushed beyond the cloud and into edge devices, why this is important, and what tools are available to do so. In summary, this paper’s contributions are a detailed review of the emerging area of LLM deployment at the edge, highlighting key applications, challenges