Edge Computing: Powering the Future of AI/ML Beyond the Cloud

Latest 50 papers on edge computing: Sep. 29, 2025

The promise of AI/ML is increasingly tied to its ability to operate where data is generated: at the edge. From autonomous vehicles to smart cities, and even space exploration, the demand for intelligent processing on resource-constrained devices is exploding. This isn’t just about faster inferences; it’s about real-time responsiveness, enhanced privacy, reduced bandwidth, and unparalleled efficiency. Recent research delves into these challenges, pushing the boundaries of what’s possible in edge computing.

The Big Idea(s) & Core Innovations

Many recent breakthroughs converge on optimizing resource utilization, enhancing security, and facilitating seamless distributed AI. A recurring theme is the intelligent allocation and management of computational and communication resources. For instance, in “SynergAI: Edge-to-Cloud Synergy for Architecture-Driven High-Performance Orchestration for AI Inference”, researchers from the National Technical University of Athens introduce SynergAI, a framework that leverages architecture-aware scheduling and QoS-driven performance optimization across heterogeneous Edge-to-Cloud systems. It dynamically schedules tasks to minimize QoS violations, achieving a 2.4x reduction compared to state-of-the-art solutions.

The challenge of efficient distributed AI extends to training, not just inference. “CollaPipe: Adaptive Segment-Optimized Pipeline Parallelism for Collaborative LLM Training in Heterogeneous Edge Networks” from the University of Example and Research Lab Inc. proposes CollaPipe, an adaptive pipeline parallelism framework that optimizes Large Language Model (LLM) training in heterogeneous edge networks. Their delay-only driven scheduling dynamically selects compute units based on local training delays, significantly improving resource utilization and training efficiency.

Security and efficiency are paramount, especially in dynamic environments. In “Robust and Secure Computation Offloading and Trajectory Optimization for Multi-UAV MEC Against Aerial Eavesdropper”, Author A and colleagues from the University X, University Y, and University Z present a framework for secure computation offloading and trajectory optimization in multi-UAV MEC systems. This ensures robustness against aerial eavesdroppers, highlighting the critical role of trajectory optimization in minimizing energy consumption and vulnerability.

Another significant area of innovation is in resource management for diverse edge environments. “KubeDSM: A Kubernetes-based Dynamic Scheduling and Migration Framework for Cloud-Assisted Edge Clusters” by Amirhossein Pashaeehir and colleagues from Amirkabir University of Technology introduces KubeDSM. This framework uses Kubernetes for dynamic scheduling and live migration, drastically reducing resource fragmentation and improving resource utilization in cloud-assisted edge clusters, outperforming default Kubernetes schedulers. This is complemented by the work in “Autonomous Task Offloading of Vehicular Edge Computing with Parallel Computation Queues” by John Doe and Jane Smith from University of Technology and Institute for Advanced Research, which proposes parallel computation queues for autonomous task offloading in vehicular edge computing, leading to significant latency reduction and throughput improvement.

Beyond traditional resource management, there’s a push for more intelligent, adaptive systems. “Multi-Agent Reinforcement Learning for Task Offloading in Wireless Edge Networks” from LIA, Avignon University, and INRIA, explores a decentralized Multi-Agent Reinforcement Learning (MARL) framework using constraint-based coordination to enable implicit coordination among agents. This reduces communication overhead and enhances scalability in large-scale edge environments. Additionally, “Constraint-Compliant Network Optimization through Large Language Models” by Youngjin Song and colleagues from Korea University demonstrates how Large Language Models (LLMs) can be leveraged for complex network optimization problems in MEC, ensuring strict constraint satisfaction through natural language-based input encoding and iterative refinement.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are often underpinned by novel architectural designs, custom accelerators, and robust evaluation methodologies:

• NVIDIA Jetson Devices: The paper “Characterizing the Performance of Accelerated Jetson Edge Devices for Training Deep Learning Models” by Prashanthi S.K. and colleagues from the Indian Institute of Science provides deep insights into optimizing DNN training on NVIDIA Jetson AGX Xavier, Xavier NX, and Nano devices, highlighting the impact of disk caching, pipelining, and power modes. Their code is available at https://github.com/dream-lab/edge-train-bench/tree/sigmetrics-2023.
• Lightweight Models & Quantization: “Comparative Analysis of Lightweight Deep Learning Models for Memory-Constrained Devices” by Tasnim Shahriar systematically evaluates MobileNetV3, ResNet18, SqueezeNet, EfficientNetV2, and ShuffleNetV2, finding MobileNetV3 offers the best balance for real-time edge applications. Complementing this, “Sensitivity-Aware Post-Training Quantization for Deep Neural Networks” by Zekang Zheng et al. introduces a sensitivity-guided post-training quantization that achieves near-lossless accuracy with significant speedups. Furthermore, “Constraint Guided Model Quantization of Neural Networks” from KU Leuven proposes CGMQ, an algorithm that automatically adjusts bit-widths for mixed-precision models under computational cost constraints.
• Specialized Hardware & Runtime Environments: “Bare-Metal RISC-V + NVDLA SoC for Efficient Deep Learning Inference” demonstrates a bare-metal implementation of RISC-V with NVIDIA Deep Learning Accelerator (NVDLA) for low-latency, high-throughput AI inference at the edge, with code available at https://github.com/LeiWang1999/ZYNQ-NVDLA. For serverless workloads, “WebAssembly and Unikernels: A Comparative Study for Serverless at the Edge” introduces Limes, a WebAssembly-based execution environment built on Wasmtime, comparing it against Firecracker and unikernels. Relatedly, “Unikernels vs. Containers: A Runtime-Level Performance Comparison for Resource-Constrained Edge Workloads” further highlights unikernels’ advantages in image size and boot time for edge workloads.
• Distributed Learning Frameworks: “CollaPipe: Adaptive Segment-Optimized Pipeline Parallelism for Collaborative LLM Training in Heterogeneous Edge Networks” offers a framework for collaborative LLM training (code: https://github.com/moon). “Multi-Worker Selection based Distributed Swarm Learning for Edge IoT with Non-i.i.d. Data” from IEEE Signal Processing Magazine and University of Toronto tackles non-i.i.d. data challenges in distributed swarm learning for edge IoT.
• AI/ML Models & Datasets for Specific Applications: “Lidar-based Tracking of Traffic Participants with Sensor Nodes in Existing Urban Infrastructure” proposes a CPU-only lidar system for real-time traffic monitoring. “Comprehensive Evaluation of CNN-Based Audio Tagging Models on Resource-Constrained Devices” evaluates PANNs and MobileNet variants on Raspberry Pi for audio tagging (code: https://github.com/gbibbo/ai4s-embedded). For BCI, “Neural Signal Compression using RAMAN tinyML Accelerator for BCI Applications” presents the RAMAN tinyML accelerator for efficient neural signal compression (code: https://github.com/raman-tinyml/BCI-Compression).

Impact & The Road Ahead

The implications of these advancements are profound. We are moving towards a future where AI isn’t confined to distant data centers but is a pervasive, intelligent layer embedded in our physical world. The transition from theoretical concepts to practical, deployable solutions is accelerating. Imagine intelligent traffic systems reacting in milliseconds, autonomous spacecraft dynamically optimizing observations to avoid clouds, or historic buildings predicting maintenance needs with federated learning, all without compromising privacy.

The rise of agentic AI, as surveyed in “Governed By Agents: A Survey On The Role Of Agentic AI In Future Computing Environments” by Nauman Ali Murad and Safia Baloch from GIK Institute, is a testament to this shift, promising to reduce reliance on large public clouds by enabling efficient, localized processing. Collaborative intelligence across cloud-edge-terminal layers, as highlighted in “A Survey on Cloud-Edge-Terminal Collaborative Intelligence in AIoT Networks” by Li Wei and colleagues, will be crucial for scalable and real-time AIoT systems. Moreover, securing these distributed systems against threats like poisoning attacks, as addressed by “Deep Learning based Moving Target Defence for Federated Learning against Poisoning Attack in MEC Systems with a 6G Wireless Model”, will remain a critical area of research.

The future of AI/ML at the edge is not just about bringing computation closer to data; it’s about making that computation smarter, more secure, more energy-efficient, and more integrated with our physical infrastructure. These papers collectively paint a picture of an intelligent, decentralized world where AI is everywhere, seamlessly enhancing our lives.