Edge Computing: The New Frontier for Intelligent, Sustainable, and Scalable AI/ML

Latest 50 papers on edge computing: Nov. 30, 2025

The world of AI/ML is increasingly decentralized, moving intelligence closer to where data is generated. Edge computing, with its promise of low-latency processing and enhanced privacy, is at the forefront of this revolution. Recent research highlights a surge in innovative solutions addressing the unique challenges and vast potential of deploying AI/ML models on resource-constrained edge devices. From optimizing large language models (LLMs) to building sustainable infrastructure and ensuring robust security, the advancements are transforming how we think about pervasive AI.

The Big Idea(s) & Core Innovations

The overarching theme uniting recent breakthroughs in edge computing is the drive for efficiency without compromise. This means achieving high performance, low latency, and energy efficiency, often under severe resource constraints, while simultaneously tackling challenges like data sparsity and dynamic environments.

One significant innovation lies in optimizing large models for edge deployment. The paper, “Towards Edge General Intelligence: Knowledge Distillation for Mobile Agentic AI” by authors from the University of Technology and MIT, emphasizes Knowledge Distillation (KD) as crucial for deploying large models efficiently on mobile and edge devices. This sentiment is echoed by “SLED: A Speculative LLM Decoding Framework for Efficient Edge Serving” from Virginia Tech and Queen’s University Belfast, which introduces a speculative decoding framework. SLED leverages lightweight draft models on edge devices with a shared target model on an edge server, achieving a remarkable ×2.2 higher system throughput and ×2.8 higher capacity without sacrificing accuracy. Similarly, “Collaborative Large Language Model Inference via Resource-Aware Parallel Speculative Decoding” proposes a resource-aware parallel speculative decoding approach to boost LLM inference efficiency.

Beyond model optimization, intelligent resource management and dynamic adaptation are critical. “Joint Edge Server Deployment and Computation Offloading: A Multi-Timescale Stochastic Programming Framework” by Zhang, Wang, and Chen from various universities, offers a stochastic programming framework to jointly optimize server deployment and computation offloading, leading to more adaptive edge resource management. “Dynamic Edge Server Selection in Time-Varying Environments: A Reliability-Aware Predictive Approach” further advances this by using predictive modeling for proactive resource management, enhancing service reliability in fluctuating network conditions. In vehicular contexts, “Energy-Efficient Task Computation at the Edge for Vehicular Services” by P. Parastar et al. proposes a mobility-aware framework that significantly reduces energy consumption, while “Reinforcement Learning for Resource Allocation in Vehicular Multi-Fog Computing” from University of X shows RL-based methods reducing latency by up to 30% in high-mobility scenarios.

Sustainable and secure edge computing is another burgeoning area. “CarbonEdge: Leveraging Mesoscale Spatial Carbon-Intensity Variations for Low Carbon Edge Computing” by Wu et al. from UMass Amherst and CMU, introduces a framework that reduces emissions by up to 78.7% by shifting workloads based on localized carbon intensity. For security, “Toward an Intrusion Detection System for a Virtualization Framework in Edge Computing” by authors from Technology Innovation Institute proposes a lightweight IDS seamlessly integrated with virtualized edge environments, balancing performance and security.

Innovative approaches also extend to specialized applications like “Edge-Based Predictive Data Reduction for Smart Agriculture: A Lightweight Approach to Efficient IoT Communication” by Fathalla et al., which uses lightweight LSTM models and adaptive transmission logic to reduce data transmission in agricultural IoT by over 90% without sacrificing accuracy. “Neuro-Inspired Task Offloading in Edge-IoT Networks Using Spiking Neural Networks” highlights the energy efficiency gains from using SNNs for task offloading, marking a shift towards bio-inspired computing at the edge.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are often underpinned by specific models, datasets, and benchmarks that push the boundaries of edge AI. Here’s a look at some notable contributions:

• Lightweight LLMs & Inference Runtimes: The paper “An Evaluation of LLMs Inference on Popular Single-board Computers” provides a comprehensive benchmark of 25 quantized open-source LLMs on SBCs like Raspberry Pi 4/5 and Orange Pi 5 Pro. It highlights the critical role of runtimes, showing Llamafile’s superior throughput and power usage over Ollama for efficient on-device AI.
• Spatiotemporal Transformers: “MoCap2Radar: A Spatiotemporal Transformer for Synthesizing Micro-Doppler Radar Signatures from Motion Capture” leverages spatiotemporal transformers to synthesize radar signatures, demonstrating efficiency and generalization across motion patterns, ideal for augmenting scarce radar datasets.
• SparseST Framework: “SparseST: Exploiting Data Sparsity in Spatiotemporal Modeling and Prediction” introduces a framework combining 2D sparse convolution with the delta network algorithm, achieving up to 90% computational savings while maintaining accuracy in video prediction and anomaly detection.
• Agentic AI Frameworks: “Towards Edge General Intelligence: Knowledge Distillation for Mobile Agentic AI” proposes hybrid architectures and outlines the core components of Agentic AI (Perception, Planning, Action, Memory) for edge-constrained environments.
• Monitoring Systems: “SERVIMON: AI-Driven Predictive Maintenance and Real-Time Monitoring for Astronomical Observatories” integrates cloud-native technologies (Prometheus, Grafana, Cassandra) with ML (Isolation Forest) for real-time anomaly detection in distributed systems.
• Specialized Algorithms: “One-Shot Knowledge Transfer for Scalable Person Re-Identification” introduces OSKT, a novel one-shot method that generates scalable person ReID models by transferring teacher model knowledge into a weight chain, outperforming traditional compression methods. Its code is available at https://github.com/SEU-CL/OSKT.
• Sustainable Edge Management: “CarbonEdge: Leveraging Mesoscale Spatial Carbon-Intensity Variations for Low Carbon Edge Computing” provides a carbon-aware placement framework with code available at https://github.com/umassos/CarbonEdge.

Impact & The Road Ahead

The implications of these advancements are profound. We’re seeing a clear trajectory towards truly intelligent and autonomous edge systems. From making AI accessible on low-cost hardware for SMEs and individual developers, to enabling critical real-time decision-making in autonomous vehicles, smart agriculture, and maritime search and rescue, the edge is becoming a powerhouse of innovation.

The emphasis on energy efficiency and sustainability (e.g., CarbonEdge, wireless-powered MEC systems like those in “Fairness-Aware Computation Offloading in Wireless-Powered MEC Systems with Cooperative Energy Recycling”) is crucial as AI adoption scales globally. The shift towards decentralized and hierarchical AI architectures provides enhanced robustness, privacy, and scalability, as demonstrated by “A Decentralized Root Cause Localization Approach for Edge Computing Environments” and “Distributed Hierarchical Machine Learning for Joint Resource Allocation and Slice Selection in In-Network Edge Systems”.

Looking ahead, the integration of cutting-edge research in neuromorphic computing (e.g., “Unsupervised local learning based on voltage-dependent synaptic plasticity for resistive and ferroelectric synapses”) hints at even more energy-efficient and biologically plausible AI at the edge. The development of robust intellectual property protection schemes like RISE in “Robust Client-Server Watermarking for Split Federated Learning” will foster trust and collaboration in decentralized AI ecosystems.

The future of AI is undeniably at the edge. These papers collectively paint a picture of a dynamic, rapidly evolving field where efficiency, intelligence, and sustainability are paramount, promising a new era of pervasive and impactful AI applications.