Intelligent Connected Vehicles (ICVs) impose stringent requirements on real-time computational services. However, limited onboard resources and the high latency of remote cloud servers restrict traditional solutions. Unmanned Aerial Vehicle (UAV)-assisted Mobile Edge Computing (MEC), which deploys computing and storage resources at the network edge, offers a promising solution. In UAV-assisted vehicular networks, jointly optimizing content and service caching, computation offloading, and UAV trajectories to maximize system performance is a critical challenge. This requires balancing system energy consumption and resource allocation fairness while maximizing cache hit rate and minimizing task latency. To this end, we introduce system efficiency as a unified metric, aiming to maximize overall system performance through joint optimization. This metric comprehensively considers cache hit rate, task computation latency, system energy consumption, and resource allocation fairness. The problem involves discrete decisions (caching, offloading) and continuous variables (UAV trajectories), exhibiting high dynamism and non-convexity, making it challenging for traditional optimization methods. Concurrently, existing multi-agent deep reinforcement learning (MADRL) methods often encounter training instability and convergence issues in such dynamic and non-stationary environments. To address these challenges, this paper proposes a MADRL-based joint optimization approach. We precisely model the problem as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) and adopt the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm, which follows the Centralized Training Decentralized Execution (CTDE) paradigm. Our method aims to maximize system efficiency by achieving a judicious balance among multiple performance metrics, such as cache hit rate, task delay, energy consumption, and fairness. Simulation results demonstrate that, compared to various representative baseline methods, the proposed MAPPO algorithm exhibits significant superiority in achieving higher cumulative rewards and an approximately 82% cache hit rate.

City-scale unmanned aerial vehicle vision-and-language navigation (UAV-VLN) requires accurate upstream target localization from an overhead map, onboard observation, and language description. Existing VLM-based methods often treat road names, landmarks, and spatial relations as raw text, leaving the model to search a large map and implicitly infer geometric constraints. This paper proposes GPOD, an inference-time candidate-prior interface for the upstream target-localization stage in city-scale UAV-VLN. GPOD converts language anchors, spatial relations, target-category cues, static map objects, and vehicle detections into ranked candidate priors through branch-specific candidate generation, thereby reformulating unconstrained full-map coordinate regression as candidate-prior-conditioned coordinate prediction. The static branch aligns language constraints with map-object geometries, while the dynamic branch uses YOLOv8l-VisDrone with Slicing Aided Hyper Inference (SAHI) to construct detection-conditioned vehicle candidates. In the GPOD-VLM setting, ranked candidates are injected as structured spatial prompts and the base VLM predicts the final continuous coordinates; GPOD-Direct is a candidate-direct diagnostic variant that directly uses candidate centers without VLM coordinate regression. On the CityNav localization protocol, GPOD improves FlightGPT Overall SR@20m from 15.23% to 25.61% and consistently reduces Mean Navigation Error (Mean NE) across splits and backbones. On Val-Unseen, GPOD-Direct (Top-1) reaches 32.59% SR@20m, showing that ranked candidate priors provide strong discrete localization signals. These results show that inference-time candidate priors can reduce city-scale search ambiguity without updating the base VLM parameters, while also revealing a candidate-utilization gap in the current prompt-based continuous coordinate-regression interface.