Aug 2026· Applied and Computational Engineering· 0 citations
Abstract
As 6G networks advance toward higher levels of autonomy and intelligence, the demand for sophisticated multimodal data processing in communication systems is growing exponentially. Conventional localized AI models encounter significant generalization bottlenecks when handling cross-layer network operations and dynamic resource allocation. To overcome these limitations, this paper systematically investigates the application frameworks of large language models (LLMs) in wireless communication systems—spanning from physical-layer protocols to high-layer network management—while critically evaluating the associated deployment challenges. Drawing on a comprehensive review of prominent literature published over the past three years, this study empirically assesses the performance of diverse LLM architectures across three key domains: physical-layer protocol parsing, network-layer resource allocation, and service orchestration. Results demonstrate that LLMs yield substantial improvements in end-to-end semantic communication, standardized protocol interpretation, and intelligent network resource scheduling. Nevertheless, practical deployment remains severely hindered by the computational constraints of edge devices and prohibitively high inference latency. We conclude that the co-design of lightweight, telecom-specific large language models (Telecom-LLMs) and distributed inference mechanisms constitutes a pivotal evolutionary pathway toward realizing endogenous intelligence in future wireless communication systems.
Decentralized large language model (LLM) inference distributes transformer layers across heterogeneous nodes to enable resource-constrained participants to collaboratively serve large models. However, inter-node activation transfer constitutes a significant communication bottleneck, especially under low-bandwidth conditions prevalent in consumer-grade and edge networks. We propose BandwidthLLM, a communication-efficient framework that integrates three techniques: (1) a bandwidth-aware layer placement algorithm that minimizes boundary-level transfer cost according to link bandwidth and node reliability; (2) a lightweight activation compression scheme combining adaptive quantization with outlier-aware clipping and error feedback; and (3) a semantic preservation check that automatically falls back to higher precision when compressed activations deviate beyond a calibrated threshold. Experiments on Llama-2-7B, Mistral-7B, and Qwen2.5-7B across simulated bandwidth conditions (20–500 Mbps) demonstrate that BandwidthLLM reduces inter-node activation traffic by 62.7–63.4% relative to the BloomBee-like baseline while maintaining perplexity degradation of at most 0.31 points and task accuracy within 0.9 percentage points of uncompressed inference. Under 20 Mbps links, BandwidthLLM achieves 1.7× higher throughput and 41% lower latency.
Zhihao Wang, Shi Shu, Yulin Wang et al.· 2026 8th International Confe...· 1 citation
Inference-time scaling has emerged as an effective approach for enhancing the capabilities of Large Language Models (LLMs), addressing the growing demand for stronger reasoning without increasing model size. This novel form of LLM scaling comprises two representative approaches: explicit reasoning, which generates intermediate chain-of-thought tokens during an explicit thinking phase, and implicit reasoning, which iteratively updates hidden states in the latent space without producing explicit outputs. Despite their effectiveness, both paradigms incur substantial computational and memory overhead, raising challenges for deployment on resource-constrained edge devices. To address these issues, we propose a Mobile Reasoning-as-aService (MORES) framework that treats reasoning as a computational service accessible to edge devices over wireless networks. Focusing on implicit reasoning, we leverage its recursive structure to partition hiddenstate updates between edge devices and servers, enabling cooperative inference that allows devices to access additional cloud computation on demand. To optimize long-term performance, we formulate a joint computation and communication scheduling problem and solve it using a semantic Mixture-of-Experts (MoE)-based Deep Reinforcement Learning (DRL) algorithm to address heterogeneity in wireless conditions and task demands. The agent adaptively allocates resources by adjusting the number of recurrent steps and the transmission pruning rate, while a semantic router enables high-speed gating for real-time expert selection. Experimental results show that the proposed method achieves an approximately 18% improvement in system throughput over the baseline Soft Actor-Critic (SAC) algorithm. Our code is available at https://github.com/NICE-HKU/MORES.
Guanchen Liu, Hongyang Du, Kaibin Huang· 1 citation
Wireless Mesh Networks (WMNs) are essential for many Internet of Things (IoT) applications, such as industrial automation, environmental monitoring, and smart cities. Today, configuring a WMN to meet its stringent performance requirements remains a significant challenge due to dynamic real-world wireless conditions and operating environments. The simulationto-reality gap in network configuration further complicates the generalization of models trained solely with simulation data, leading to suboptimal performance in physical deployments. To address such challenges, we develop WMN-LLM-DS, a novel framework that integrates Large Language Models (LLMs) for synthetic data generation with domain adaptation techniques to better configure WMNs. Leveraging LLMs, WMN-LLM-DS generates high-quality, diverse synthetic datasets conditioned on realworld constraints, effectively bridging the simulation-to-reality gap and enriching the diversity of training data. WMN-LLM-DS employs a teacher-student architecture to transfer network configuration knowledge learned from simulations to physical deployments, enhanced by custom loss functions to align feature representations across different domains. Experiments conducted on datasets collected from a physical testbed and network simulators demonstrate that WMN-LLM-DS outperforms existing solutions, achieving an improvement of up to 10.8% in prediction accuracy while also exhibiting strong domain generalization capabilities.
Aitian Ma, Jean Marco Cruz, Dongsheng Luo et al.· Annual International Compute...· 1 citation
Large Language Model (LLM) serving infrastructure has evolved from single-model deployments into heterogeneous fleets combining general-purpose models, domain-specialised variants, multimodal models, and quantised derivatives with widely differing context capacities. In such environments, the decision of which endpoint should serve a given request materially affects latency, throughput, monetary cost, and output quality. Despite growing industrial adoption of inference routers, the literature lacks a consolidated account of the decision dimensions such systems must reason over, or of how those dimensions interact when their objectives conflict. This paper addresses that gap. Through an analysis of inference-serving mechanisms and model-specialisation results, we identify and characterise six dimensions governing endpoint selection: load distribution, context-length requirement, input modality, task category, domain specialisation, and prefix-cache locality. For each dimension we formalise the decision criterion and describe the signals available at request-admission time, distinguishing hard feasibility constraints from soft optimisation preferences. We further show that these dimensions are not independent - notably, cache-locality routing and load-balancing objectives are structurally opposed, since the former concentrates traffic while the latter disperses it - and we propose a bounded-affinity policy and a priority-ordered evaluation sequence to resolve such conflicts. Finally, we specify an evaluation protocol comprising workload definitions, baselines, and metrics, by which implementations of the framework may be empirically validated and compared. The framework is intended as a design reference for practitioners building inference gateways and as a structuring basis for future empirical work on multi-objective LLM request scheduling.
Rajalakshmi Srinivasaraghavan· International journal of com...· 0 citations
As Software-Defined Networking (SDN) and Network Function Virtualization (NFV) enabled networks scale in size and complexity, monitoring and managing Service Function Chains (SFCs) under stringent latency and resource constraints becomes increasingly challenging. Although Deep Reinforcement Learning (DRL) is widely applied to SFC provisioning and Virtual Network Function (VNF) placement, enhanced network state monitoring is crucial to capture unexpected network conditions and guide DRL agents toward more adaptive decisions. In this context, Language Models (LMs) enable flexible, natural-language (NL)–based, query-driven network monitoring; however, directly processing complex multi-metric NL queries is computationally expensive and error-prone. This paper proposes an end-to-end (E2E) edge-based query translation pipeline that decomposes multi-metric NL queries into simpler single-metric sub-queries. Query decomposition is performed using a retrieval-augmented language model (RAG-LLM) and compared with a lightweight rule-based decomposition baseline. The resulting sub-queries are translated into Structured Query Language (SQL) using FLAN-T5. A cloud-only baseline, which directly translates NL queries to SQL without decomposition, is also evaluated. The results show that the rule-based edge pipeline achieves the lowest latency, reducing E2E latency by up to 78% compared to RAG-LLM and 18% compared to cloud execution under high workloads. Under increasing arrival rates for the largest workload, the rule-based edge pipeline maintains superior performance over cloud, reducing total E2E latency by 57% at $\lambda = 0.8$ . While RAG-LLM provides greater flexibility for unseen query patterns, both edge-based approaches achieve 100% NL2SQL accuracy with zero decomposition failures, outperforming the cloud-only baseline (95% accuracy).
Parisa Fard Moshiri, Xinyu Zhu, Poonam Lohan et al.· IEEE Transactions on Network...· 0 citations
Efficient long-term network evolution is becoming increasingly critical in dense 5G-Advanced and beyond cellular systems, where persistent traffic imbalances and localized congestion pose significant challenges that conventional short-term radio resource management alone cannot fully mitigate. This paper proposes a digital twin (DT)-enabled non-real-time (NRT) network evolution framework integrated with a large language model (LLM). Within this architecture, the digital twin provides a high-fidelity, controllable environment for evaluating infrastructure actions, while the LLM serves as a strategic orchestration engine that recommends cost-efficient network upgrades based on observed network states. Unlike traditional optimization methods that require exhaustive mathematical reformulations for each specific scenario, the proposed framework leverages the reasoning capabilities of LLMs to interpret operator objectives and constraints in natural language, generating structured evolution plans. The considered NRT action space encompasses antenna upgrades, bandwidth expansion, and new base station (BS) deployment. A techno-economic formulation is introduced to jointly evaluate load reduction performance and overall economic expenditure. Numerical results in a dense cellular scenario demonstrate that the framework effectively reduces peak resource utilization and provides diverse, coordinated evolution strategies tailored to varying network conditions.
Yukai Wang, Janghee Woo, G. Hahm et al.· International Conference on...· 0 citations