The growing demand for edge intelligent applications has driven deployment of Large Language Models (LLMs) on personal edge devices, where Mixture-of-Experts (MoE) enable LLMs to scale parameter capacity through sparse expert activation. To further accelerate model inference, speculative decoding (SD) is a promising technique that verifies multiple candidate tokens in parallel. However, deploying MoE on resource-constrained edge devices relies on expert offloading due to limited GPU memory, and SD-induced parallel verification activates a large set of experts, amplifying CPU–GPU I/O overhead and diminishing the expected speedup. To address this issue, we propose Liam, a Locality-enhanced inference acceleration method for MoE models on personal edge devices with consumer-grade GPUs. Liam is motivated by the insight that expert activations during speculative verification exhibit spatiotemporal locality. Specifically, a dual-stage decoding is designed by an approximate-then-correct manner. First, the fast approximation stage runs for multiple rounds by statistically maintained expert caches to enable expert substitution based on routing score distance. Followed by periodic correction stage with feedback-driven speculation length control, Liam realizes the trade-off between accepted tokens and speedup. Finally, extensive experiments on resource-constrained personal edge platforms demonstrate that Liam effectively improves inference throughput compared to existing baselines while maintaining energy efficiency.
Vehicular CrowdSensing (VCS) represents a unique subset of crowdsourcing, where task assignment remains a fundamental and challenging issue. However, traditional task assignment methodologies rarely consider how task assignments are accomplished under chance constraints induced by practical uncertainties. In this paper, we investigate the Robust Task Assignment problem for Probabilistic constraints (RTA-P), considering nondeterministic vehicle arrivals and stochastic execution abilities. To ensure the quality of task completion, it is necessary to recruit a sufficient number of vehicles, thereby enhancing task robustness. The RTA-P, constrained by joint probabilities, is identified as a non-submodular problem. By leveraging the submodular characteristics of Poisson Binomial Distributions (PBDs), we transform this into a Minimum Submodular Cover with a Submodular Cost (MSC/SC) problem. We then consider the problem in both offline and online scenarios. In the offline scenario, we employ a greedy algorithm and analyze its time complexity. We further introduce an improved time-efficient algorithm and analyze the approximation ratios of both algorithms. In the online scenario, the task completion probability of vehicles varies stochastically across different rounds, following an unknown distribution. We tackle this online stochastic problem within an online learning framework based on the optimistic principle in the face of uncertainty, utilizing offline algorithms as oracles to guide decision-making. Finally, we evaluate the proposed algorithms using one synthetic dataset and three real datasets. The experimental results demonstrate the effectiveness, efficiency, and scalability of our proposed algorithms.
Unmanned aerial vehicle (UAV)-enabled mobile edge computing (MEC) can extend onboard computing capability through wireless task offloading. For privacy-sensitive aerial services, however, even the occurrence and timing of offloading can expose service activity to unauthorized monitoring, while practical airspace restrictions limit mobility for mitigating such exposure. This paper investigates cross-boundary covert task offloading, where a task-generating UAV outside a polygonal no fly zone (NFZ) offloads to multiple edge servers (ESs) inside it under surveillance by a ground warden Willie. We maximize completed offloaded workload by jointly optimizing the UAV trajectory, ES scheduling, and spectral efficiency subject to flight, task-arrival, edge-execution, and mission-level covertness con straints. To characterize rare but consequential covert failures, we derive the per-slot covert-outage probability over Nakagami-m fading and define mission-level outage as at least one covertness violation, whose probability constraint is exactly transformed into an additive covert-risk budget. To tackle the resulting mixed discrete-continuous non-convex problem, we develop a two-stage framework that constructs homotopy-guided convex corridors and performs joint offloading and trajectory optimization under covert-risk and computation constraints. Numerical results show that the proposed framework outperforms designed benchmarks. More importantly, we reveal that, in such risk-aware covert regimes, stronger fading fluctuations increase the likelihood of detection-favorable extreme realizations and are therefore less favorable for covert communication.
Monitoring points of interest (PoIs) representing water bodies and related infrastructures in hydraulic metaverses, with the help of unmanned aerial vehicles (UAVs), is central to protecting aquatic ecosystems and ensuring public safety. However, compared to traditional snapshot inspections, continual monitoring relies on UAV swarms to collect data from PoIs in a persistent manner, which must navigate complex operational constraints including limited battery capacity, strict user delay requirements, and dynamic weather conditions, while ensuring that learning-based detection models remain robust over time without suffering from catastrophic forgetting. To address this challenge, we propose an automatic hydraulic monitoring system across multiple task scenarios, utilizing a UAV swarm orchestrated by large language model (LLM) agents equipped with a suite of specialized tools. First, an environment perception tool is designed for the agent to observe dynamic global information, including regional weather and metaverse network topology. Building on this, a self-evolving swarm scheduling tool is developed to collect data from PoIs, where a feedback-based refinement module initially predicts the required number of UAVs and iteratively improves its decisions by leveraging verified feedback from a trajectory planning module operating in environments. To accommodate different learning paradigms across application scenarios, a dualmode model update tool is further proposed to support both snapshot training with one-time full-data collection and continual learning through progressive data stream acquisition. Finally, experimental results demonstrate that our system outperforms baseline methods across multiple performance metrics, offering a continual monitoring system for the hydraulic metaverse.
Large Language Models (LLMs) have become foundational to modern AI, transforming how we interact with devices, access information, and enabling the development of powerful, intuitive virtual assistants. While current LLM training is confined to cloud data centers due to massive computational and data demands, fine-tuning these models on resource-constrained devices such as consumer smartphones remains largely unexplored. In this paper, we present Confidant, a practical and collaborative training framework that enables fine-tuning of modern LLMs across multiple off-the-shelf mobile devices. Confidant partitions an LLM into smaller sub-models that individually fit within the memory limits of mobile hardware, and orchestrates their collaborative training through a novel pipeline-parallel approach. It incorporates a memory-aware dynamic model partitioning strategy and an intra-device multi-processor scheduler to minimize training latency across heterogeneous platforms. To ensure robustness, Confidant introduces a hybrid fault-tolerance mechanism that proactively addresses device and network failures. To further mitigate communication bottlenecks under low-bandwidth or unstable networks, Confidant incorporates a lightweight Dynamic Communication Controller that adaptively reduces transmission volume to minimize communication delay. We implement Confidant in C++/Python, including a cross-framework adapter to support diverse mobile environments. Experimental results show that Confidant supports the fine-tuning of models across a wide range of scales on heterogeneous edge devices, ranging from hundred-million-parameter models such as BERT-Base and GPT-2 Medium to billion-parameter models such as Phi2 and LLaMA3. Using Confidant, we are able to finetune the 2.7-billion-parameter Phi-2 model on the Alpaca dataset in 40.1 hours using one smartphone and two laptops.
