Open research questions in IoT and Edge/Fog Computing

What the literature leaves open

• Further research is needed to improve the accuracy of the Digital Twin - Further research is needed to apply the Digital Twin to other domains - Further research is needed to investigate the effects of other factors on LLM-adapter serving

  Data-driven optimization of GPU efficiency for distributed LLM-adapter serving · 2026 · DOI

• existing approaches focus on proactive placement strategies - no existing simulator or digital twin explicitly models the combined dynamics of adapter caching, KV-cache allocation, and continuous batching

  Data-driven optimization of GPU efficiency for distributed LLM-adapter serving · 2026 · DOI

• One challenge is the resource constraints in TinyML environments, which require significant adaptations of machine learning algorithms. Another challenge is the need to balance performance metrics such as model accuracy, inference latency, memory utilization, and power consumption in TinyML systems. The paper also highlights the challenge of latency in TinyML solutions, which must account for not only the inference computational time but also delays in preprocessing steps and memory access operations.

  Tiny Machine Learning and On-Device Inference: A Survey of Applications, Challenges, and Future Directions · 2025 · DOI

• Significant heterogeneity in the TinyML hardware ecosystem, - Low standardization in TinyML hardware, - Device specifications depend on the use case, - Limited by the constraints of edge devices

  Tiny Machine Learning and On-Device Inference: A Survey of Applications, Challenges, and Future Directions · 2025 · DOI

• Existing surveys focus on traditional centralized offloading approaches or emphasize reinforcement learning with limited integration of deep learning. There is a lack of comprehensive and focused surveys on the full-scale application of DRL to the task offloading problem in fog computing environments. Unexplored areas and new directions for advancing DRL-based solutions in fog computing need to be identified.

  Analysis of Deep Reinforcement Learning Algorithms for Task Offloading and Resource Allocation in Fog Computing Environments · 2025 · DOI

• Radio traffic is far costlier than sensing or computing, dominantly draining batteries. Protocol choice is central, with different protocols differing in header size, state complexity, and security load. The lack of a unified evaluation framework hinders the development of comparable quantitative metrics across protocols.

  Energy Footprint and Reliability of IoT Communication Protocols for Remote Sensor Networks · 2025 · DOI

• Literature remains siloed, focusing on single verticals or Layer-2 aspects, - Quantitative metrics are lacking across protocols, - Direct quantitative benchmarking is problematic due to environmental and hardware factors

  Energy Footprint and Reliability of IoT Communication Protocols for Remote Sensor Networks · 2025 · DOI

• Investigating the algorithm's performance in real-world scenarios - Exploring the application of the algorithm in other edge computing systems - Evaluating the algorithm's scalability and robustness

  Multi-Agent Deep Reinforcement Learning Based Dynamic Task Offloading in a Device-to-Device Mobile-Edge Computing Network to Minimize Average Task Delay with Deadline Constraints · 2024 · DOI

• The existing literature has insufficiently explored the dynamic partitioning of idle and active devices in D2D-MEC systems. The challenges of large action space and the coupling of actions across time slots have not been adequately addressed. The paper identifies the need for a novel algorithm to minimize the long-term average delay of delay-sensitive tasks under deadline constraints.

  Multi-Agent Deep Reinforcement Learning Based Dynamic Task Offloading in a Device-to-Device Mobile-Edge Computing Network to Minimize Average Task Delay with Deadline Constraints · 2024 · DOI

• We also introduce a taxonomy of DRL-based task offloading models and highlight key challenges, open issues, and future research directions.

  Analysis of Deep Reinforcement Learning Algorithms for Task Offloading and Resource Allocation in Fog Computing Environments · 2025 · DOI

• Furthermore, we identify and analyze the open challenges and research gaps that remain to be addressed, including data gathering, scalability, interpretability, robustness, and ethical considerations.

  Cognition and context-aware decision-making systems for a sustainable planet: a survey on recent advancements, applications and open challenges · 2025 · DOI

• Ethical and privacy issues dictated that real patient data would not be used; instead, synthetic heat-stroke healthcare data consisting of 10,000 controlled records obtained through anonymized randomization was applied.

  ImmutableShield: A Fragmentation-Driven Security Algorithm for Real-Time Health Data Protection in IoT-Enabled Heat-Stroke Scenarios · 2026 · DOI

• Managing heterogeneous data traffic is a crucial challenge in IoT networks. There is a need for a dynamic channel allocation method that can support both real-time and non-real-time data transmission.

  Design and Performance Analysis of a Dynamic Channel Allocation Method for IoT Systems with Finite Number of Nodes and Heterogeneous Data Traffics · 2026 · DOI

• Many existing studies emphasize architectural descriptions without addressing decision latency, operational trade-offs, or performance validation. The gap in existing research is the lack of a formal decision placement framework supported by quantitative evaluation.

  Exploring the Future of IoT and its Integration with Cloud Computing · 2026 · DOI

• Existing deep reinforcement learning approaches have limitations in non-stationary environments. Conventional deep reinforcement learning approaches incur high computational overhead.

  An adaptive deep reinforcement learning framework enhanced by broad reinforcement learning–based state transition for efficient task offloading and resource allocation at the network edge · 2026 · DOI

• The scalability of BRL-STOA is evaluated only for state dimensions ranging from 10 to 50; performance evaluation with significantly larger-scale edge networks and thousands of mobile devices is needed.

  An adaptive deep reinforcement learning framework enhanced by broad reinforcement learning–based state transition for efficient task offloading and resource allocation at the network edge · 2026 · DOI

• Hardware constraints, model compression requirements, and system scalability remain significant. System performance is influenced by hardware limitations, energy constraints, and model optimization requirements. Scalability remains a key consideration, particularly when coordinating multiple edge nodes in distributed environments.

  Edge AI: Embedded Intelligence for Decentralized Data Analytics · 2026 · DOI

• Traditional cloud-centric analytics frameworks introduce latency, bandwidth congestion, and privacy vulnerabilities. The gap between the need for real-time data analytics and the limitations of traditional cloud-based architectures.

  Edge AI: Embedded Intelligence for Decentralized Data Analytics · 2026 · DOI

• The lack of mechanisms to address data privacy concerns in IoT architecture. The need for a lightweight and efficient approach to deploy reliable models on edge devices.

  Addressing Data Privacy Concerns in IoT Architecture with Federated Learning and TinyML · 2026 · DOI

• Existing cloud-based and centralized learning approaches introduce high communication overhead and latency. Traditional distributed learning solutions struggle with heterogeneous device capacities and model degradation.

  Latency-efficient edge intelligence in IoT networks using knowledge distillation · 2026 · DOI

• GPU utilization analysis (Figure 8) compares four methods but does not investigate how FEKD performs on edge devices without GPU acceleration or with heterogeneous compute resources (TPU, NPU, CPU-only inference) prevalent in constrained IoT edge nodes, leaving a gap in understanding applicability to truly resource-limited deployments.

  Latency-efficient edge intelligence in IoT networks using knowledge distillation · 2026 · DOI

• Existing task deployment frameworks optimize at the model level, lacking fine-grained resource awareness and concurrency control. The deployment of deep learning tasks in heterogeneous edge GPU clusters is challenging due to resource constraints and real-time requirements.

  Coconut: Multilevel Collaborative Deployment for Real-Time Deep Learning Tasks in Heterogeneous Edge GPU Cluster · 2026 · DOI

• The optimization ideas across model layer and layer level in this article can be directly applied, but the introduction of dynamic tensor shapes [86] makes the computational load of each operator highly uncertain, so more advanced operator scheduling within the stage should be studied [87], [88], which is our future work.

  Coconut: Multilevel Collaborative Deployment for Real-Time Deep Learning Tasks in Heterogeneous Edge GPU Cluster · 2026 · DOI

• Existing works primarily focus on end-to-end latency and neglect critical factors such as server operational costs and task arrival dynamics. The paper aims to bridge this gap by formulating the online scheduling problem as a mixed integer program and decomposing it into subproblems.

  Time-Dependent Path Selection and Online Learning for Efficient DAG Task Offloading in In-Network Computing · 2026 · DOI

• The paper identifies a gap in the current approach to on-device AI, which relies on compressing large transformer-based models to fit edge hardware constraints. The paper argues that this approach has limitations in terms of reasoning capabilities. The paper identifies a need for a new approach that can provide qualitatively superior intelligence, including interventional and counterfactual reasoning.

  Causal Topologies for Edge Computing: Intrinsic Efficiency and Deep Intelligence Beyond the Transformer Paradigm · 2026 · DOI

Most-cited papers in IoT and Edge/Fog Computing

• The Emergence of Edge Computing · Computer · 2017 · 2,460 citations
• On Multi-Access Edge Computing: A Survey of the Emerging 5G Network Edge Cloud Architecture and Orchestration · IEEE Communications Surveys & Tutorials · 2017 · 1,617 citations
• The Promise of Edge Computing · Computer · 2016 · 1,081 citations
• Deep Reinforcement Learning for Online Computation Offloading in Wireless Powered Mobile-Edge Computing Networks · IEEE Transactions on Mobile Computing · 2019 · 1,037 citations
• Fog Computing: Helping the Internet of Things Realize Its Potential · Computer · 2016 · 891 citations
• Edge AI: On-Demand Accelerating Deep Neural Network Inference via Edge Computing · IEEE Transactions on Wireless Communications · 2019 · 859 citations
• AI for next generation computing: Emerging trends and future directions · Internet of Things · 2022 · 585 citations
• Consumer and Object Experience in the Internet of Things: An Assemblage Theory Approach · Journal of Consumer Research · 2017 · 581 citations
• Edge AI: A survey · Internet of Things and Cyber-Physical Systems · 2023 · 551 citations
• Deep Reinforcement Learning for Task Offloading in Mobile Edge Computing Systems · IEEE Transactions on Mobile Computing · 2020 · 520 citations

Most recent work

• Energy, Scalability, Data, and Security in Massive IoT: Current Landscape and Future Directions · IEEE Internet of Things Journal · 2026
• Latency Reduction in Immersive Systems through Request Scheduling with Digital Twin Networks in Collaborative Edge Computing · ACM Transactions on Sensor Networks · 2026
• Paradigm Shift Toward Distributed Learning in IoT Intelligence: A Comprehensive Survey of Opportunities and Challenges · IEEE Internet of Things Journal · 2026
• Toward Smart 5G and 6G: Standardization of AI-Native Network Architectures and Semantic Communication Protocols · IEEE Communications Standards Magazine · 2026
• Cross-Domain Standardization and Secure Edge Intelligence for Real-Time Digital Twin Deployments in Next-Generation Communication Systems · IEEE Communications Standards Magazine · 2026
• IoT Service Orchestration in Edge–Cloud Continuum With 6G: A Review · IEEE Internet of Things Journal · 2026
• Adaptive Federated Learning for Future IoV-Oriented IoT End-to-End Network Planning · IEEE Internet of Things Journal · 2026
• D3QN-LMA: A memory-augmented deep reinforcement learning framework for energy-latency tradeoff optimization in mobile edge computing · Advanced Engineering Informatics · 2026
• c2mec: Cooperative Multi-split and Multi-hop Edge Computing Based on Deep Reinforcement Learning · ACM Transactions on Embedded Computing Systems · 2026
• Collaborative Inference and Learning between Edge SLMs and Cloud LLMs: A Survey of Algorithms, Execution, and Open Challenges · ACM Computing Surveys · 2026

Related topics in Computer Science