Agentic AI systems are reshaping communications and networking by deploying autonomous intelligent agents capable of collaborative learning while maintaining data privacy at network edges. Within distributed network environments, Multimodal Large Language Models (MLLMs) serve as cognitive engines for edge devices, yet federated fine-tuning faces substantial challenges in balancing global knowledge aggregation with local adaptation under heterogeneous network conditions. Conventional federated protocols typically rely on uniform parameter aggregation, which conflates domain-invariant features with client-specific nuances, thereby resulting in suboptimal personalization and excessive communication overhead. To address these challenges, we propose PFAdapter, a communication-efficient framework introducing hierarchical LoRA decomposition to explicitly separate adapter parameters into global-shared and local-private components. Query and key projections are assigned to global synchronization for capturing universal multimodal semantics across the network, while value and output projections remain localized for edge-specific adaptation. Additionally, orthogonality regularization based on the Frobenius norm enforces strict separation between these components, preventing redundant feature learning. Selective aggregation protocols synchronize only global-shared components across the federated network, preserving local expertise and reducing communication costs by nearly 50%. Extensive experiments on VQA-RAD, SLAKE, Hateful Memes, and CrisisMMD datasets demonstrate that PFAdapter consistently outperforms state-of-the-art baselines, achieving accuracy improvements ranging from 2.4% to 4.8% across diverse edge intelligence tasks. Consequently, our framework establishes an efficient solution for agentic AI deployment in resource-constrained communication networks.
Achieving accurate and scalable traffic prediction is a cornerstone of modern intelligent transportation systems. While the recent advent of Large Language Models (LLMs) has introduced remarkable zero-shot capabilities to urban computing, their immense parametric scale and exorbitant deployment costs fundamentally restrict their viability for real-time, large-scale traffic operations. Concurrently, mainstream spatio-temporal forecasting methods remain heavily tethered to explicit topological graph structures or computationally expensive attention mechanisms, which inherently introduce severe scalability bottlenecks. These limitations underscore a critical divergence in current research: an urgent necessity to develop highly efficient, lightweight predictive architectures capable of bridging the gap between theoretical accuracy and practical deployability.To address this deployment imperative, we propose the Spatio-Temporal Multi-view MLP Network (STMMN), a highly scalable, graph-free computational framework. Diverging from traditional graph-dependent paradigms, STMMN utilizes an adaptive spatial clustering multilayer perceptron (MLP) module to autonomously extract latent spatial homogeneity. This mechanism streamlines spatial dimensionality, effectively circumventing the rigidity of predefined topologies and drastically reducing learning complexity. Furthermore, we introduce a temporal multi-view aggregation module that seamlessly integrates distinct temporal branches, specifically capturing proximity, trend, and periodicity, to dynamically fuse multi-scale temporal dependencies without relying on exhaustive attention maps.Extensive experiments conducted on multiple real-world, large-scale datasets confirm that STMMN achieves state-of-the-art prediction accuracy while exhibiting exceptional computational efficiency. Ultimately, this research validates the immense practical value of lightweight, knowledge-aware models in overcoming the deployment bottlenecks of large-scale urban engineering operations.
