2026· IEEE Transactions on Network and Service Management· Vol 23, pp. 5980-5994· 0 citations· 45 references
Computer Science
Abstract
Large Language Models (LLMs) have been increasingly applied to intelligent vehicular systems for tasks such as scene understanding, intent reasoning, and natural language interaction. However, their inference demands exceed onboard processing capabilities, making low-latency on-vehicle inference impractical. Although edge computing can partially offload computation, the prolonged nature of LLM inference often causes execution to exceed the residence time of vehicles within edge coverage areas, leading to frequent service interruption. To address these challenges, we propose a collaborative spatiotemporal resource orchestration architecture for LLM inference in vehicular-edge networks (CoInfer). CoInfer exploits the intrinsic decomposability of LLM inference by modeling each request as a Directed Acyclic Graph (DAG) of interdependent subtasks, which are then scheduled, migrated, and aggregated along the road network to preserve end-to-end inference continuity. To improve latency and resource efficiency, CoInfer integrates multi-agent reinforcement learning for coarse-grained task orchestration with a reactive scheduler for fine-grained resource adaptation, forming a closed-loop service optimization under dynamic resource conditions. The simulation results demonstrate that CoInfer achieves a task success ratio of up to 96.0% and reduces the end-to-end inference latency by 35.7% compared to representative baselines.
Vehicular edge computing (VEC) has emerged as a key paradigm to support computation-intensive and delay-sensitive vehicular applications by offloading tasks from vehicles to nearby multi-access edge computing (MEC) servers. However, in realistic urban environments, task processing performance is heavily affected by heterogeneous vehicle-MEC interactions, spatiotemporal traffic dynamics, and continuously varying vehicle populations. To address these challenges, this paper considers a traffic-aware embodied edge intelligence-enabled vehicular network (EEIVN), where edge intelligence is grounded in the physical traffic environment by integrating VLM-based semantic perception with edge decision making. Based on this architecture, we formulate a reliability-constrained delay minimization problem (RDMP) by jointly optimizing task offloading ratio, computing resource allocation, and vehicle association, while constraining the queue reliability to mitigate queue-induced tail delay. To solve the NP-hard RDMP, we propose a VLM-multi-agent proximal policy optimization (VLM-MAPPO) approach that integrates a VLM-based traffic awareness method, a vehicle-adaptive MAPPO algorithm, and a vehicle association scoring and selection mechanism. Extensive simulations based on SUMO and CARLA demonstrate that the proposed VLM-MAPPO approach outperforms benchmarks in terms of task completion delay and tail delay, while maintaining comparable vehicle energy consumption and exhibiting robust scalability under dynamic traffic conditions and varying vehicle densities.
Xulong Qiao, Jian Wang, Zemin Sun et al.· IEEE Transactions on Cogniti...· 0 citations
Multi-agent systems (MAS) coordinate multiple LLM-powered agents through structured workflows, gaining reasoning power but incurring high inference latency from multi-step execution and repeated model invocations. Existing orchestration methods primarily optimize task performance and inference cost, leaving latency largely unaddressed. In MAS, end-to-end latency is governed by the \textit{critical execution path}, so reducing total cost alone does not reliably reduce latency. Moreover, optimizing latency while preserving accuracy remains non-trivial: naive latency optimization can misassign operator-level credit and degrade task accuracy. To address this gap, we propose \textbf{L}atency-\textbf{A}ware \textbf{M}ulti-\textbf{a}gent \textbf{S}ystem (\textbf{LAMaS}), a latency-aware orchestration framework for learning-based multi-agent systems. LAMaS addresses this challenge at two levels: at \emph{training time}, it learns latency-aware execution graphs through constrained optimization with critical-path-aware credit assignment; at \emph{inference time}, since a graph committed at training time cannot exploit runtime evidence, it complements graph construction with a lightweight controller that adaptively eliminates redundant future agent interactions as execution unfolds. Experiments on four benchmarks show that LAMaS achieves the best latency among evaluated learning-based MAS baselines, reducing end-to-end latency by over 50% while maintaining competitive or better accuracy. LAMaS is also modular and transfers to other MAS with minimal changes, consistently yielding latency reductions.
Network-wide coordinated Traffic Signal Control (TSC) is critical for enhancing urban mobility. However, existing approaches face a fundamental trade-off: traditional Multi-Agent Reinforcement Learning (MARL) is often hindered by a myopic observational scope, while Large Language Model (LLM) agents are constrained by high inference costs and spatial-topological hallucinations. To address these limitations, we present Astra, an Asynchronous Synergistic Traffic Regulation Architecture that decouples high-level strategic reasoning from reactive execution through a three-layer hierarchy. Specifically, the Macro layer performs low-frequency strategic inference to identify global bottlenecks, which are then partitioned into regional congestion subgraphs. The Meso layer employs a topological semantic causal logic mechanism to ground LLM-based coordination in physical reality via semantic primitives. Simultaneously, the Micro layer governs the broader network using decentralized MARL agents optimized with spatial attention and auxiliary prediction for robust, high-frequency execution. To bridge the gap between reasoning depth and real-time constraints, Astra incorporates an asynchronous synergetic protocol featuring strategic locking for stability and proactive feedback for adaptive re-planning. Extensive evaluations on real-world datasets demonstrate that Astra consistently outperforms state-of-the-art baseline methods, offering superior efficiency, robustness, and generalization across diverse urban scenarios.
Fansheng Sun, Jiyu Wang, Zhidan Liu· Proceedings of the 32nd ACM...· 0 citations
This paper employs five lightweight machine learning models to predict query-specific latency, accuracy, resource usage, and response size for each model-size-quantization-device combination, and design a lightweight heuristic that approximates the BILP solution.
Reza Farahani, Zoha Azimi, Mario Colosi et al.· 0 citations
Leveraging on the task planning and solving capability of pretrained Large Language Models (LLMs), deploying LLM agents on Mobile Edge Computing (MEC) edge servers brings significant benefits for an Internet of Things (IoT) network for providing enhanced AI intelligence with acceptable delay. In this work, we consider the edge LLMs deployment strategy in an end-edge-cloud LLM agents system for the IoT services, which jointly determines the locations and number of LLM initializations and user requests offloading strategy in a dynamic network environment with stochastic user requests. We formulate this joint LLM Deployment and inference Tasks Offloading (LLMDTO) problem. Typically, we design an LLM service performance evaluation mechanism by measuring its processing delay with stochastic user requests arrivals by Stochastic Network Calculus (SNC). Due to the complexity of the LLMDTO problem, we decompose this joint optimization problem into two subproblems and propose an algorithm based on Multi Agent Deep Reinforcement Learning (MADRL) scheme. To accelerate the training process of the DRL, a reward model is designed by applying the Kolmogorov Arnold Networks (KAN) to return a fast reward estimation. Finally, we validate the proposed algorithm through extensive simulations and results show the effectiveness of the proposition on lower deployment cost and delay in a dynamic network environment.
Jiayi Liu, Jinshuo Wang, Yizhi Huang et al.· IEEE Transactions on Network...· 0 citations
The rapid proliferation of Internet of Things (IoT) devices has placed unprecedented pressure on the network edge, where applications such as augmented reality, real-time analytics, and autonomous navigation demand low latency and tight energy budgets that traditional cloud-centric architectures cannot meet. Multi-access Edge Computing (MEC) addresses this gap by relocating computation closer to end users, but the core question of where and how each task should be executed remains open: rulebased and single-objective offloading strategies fail to simultaneously balance service latency, energy efficiency, and user experience under dynamic, large-scale conditions. In this paper we propose TARLOT (Two-Agent Reinforcement Learning Offloading Tasks), a cooperative framework for threetier IoT–MEC–Cloud environments. TARLOT decouples the offloading decision from the resourceallocation problem and assigns each to a dedicated Q-learning agent, so that the two subproblems are specialised independently while still being optimised jointly. The framework is evaluated on PureEdgeSim under heterogeneous IoT workloads, device densities ranging from 200 to 2,400, and diverse application profiles, and is compared against five widely-used baselines (Random, Round-Robin, Trade-Off, Pure-Edge, and Pure-Cloud). At 2,400 devices, TARLOT delivers an average service time of 1.1 s (against 4.3 s for Pure-Cloud), a Quality of Experience of 0.77 (against 0.22 for Pure-Cloud), a task-failure rate below 2 % (against nearly 14 % for Pure-Cloud), and a per-device energy consumption of only 3.6 W (against 11.2 W for Pure-Cloud) — roughly a 68 % reduction. Balanced CPU utilisation across the local, edge, and cloud tiers further confirms that TARLOT prevents resource bottlenecks, establishing it as a practical solution for next-generation large-scale IoT deployments.
Oussama Lagnfdi, Marouane Myyara, A. Darif· International journal of Com...· 0 citations