PFAdapter Boosts Personalized Federated MLLMs with Hierarchical LoRA

Agentic AI systems are transforming communications by enabling autonomous intelligent agents to learn collaboratively while preserving data privacy at the network edge. However, fine-tuning Multimodal Large Language Models (MLLMs) in federated environments faces challenges in balancing global knowledge aggregation with local adaptation, especially under heterogeneous network conditions. Traditional federated protocols often aggregate parameters uniformly, which can lead to suboptimal personalization and high communication costs. To address these issues, PFAdapter proposes a communication-efficient framework that employs hierarchical LoRA decomposition. This method explicitly divides adapter parameters into global-shared components, which capture universal multimodal semantics and are synchronized across the network, and local-private components, which remain localized for edge-specific adaptation. Orthogonality regularization further ensures strict separation, preventing redundant feature learning. By selectively aggregating only global-shared components, PFAdapter reduces communication costs by nearly 50% and achieves significant accuracy improvements (2.4% to 4.8%) across various edge intelligence tasks, making it an efficient solution for agentic AI deployment in resource-constrained networks.