PFAdapter cuts communication costs 50% with hierarchical LoRA for federated AI

Federated fine-tuning of Multimodal Large Language Models (MLLMs) on edge devices faces a fundamental tension: how to share global knowledge while preserving local adaptation. Conventional federated protocols apply uniform parameter aggregation, mixing domain-invariant features with client-specific nuances, leading to suboptimal personalization and high communication overhead. To address this, a team of researchers introduces PFAdapter (arXiv:2607.12111), a communication-efficient framework that uses hierarchical LoRA decomposition. By assigning query and key projections to global synchronization (capturing universal multimodal semantics) and keeping value and output projections local (for edge-specific adaptation), PFAdapter explicitly separates shared and private components. An orthogonality regularization based on Frobenius norm enforces strict separation, preventing redundant feature learning.

Selective aggregation synchronizes only global-shared components across the federated network, preserving local expertise and reducing communication costs by nearly 50%. Extensive experiments across four datasets—VQA-RAD (medical VQA), SLAKE (medical visual QA), Hateful Memes (multimodal hate speech), and CrisisMMD (disaster response)—show PFAdapter consistently outperforms state-of-the-art baselines, achieving accuracy gains of 2.4% to 4.8%. The framework establishes an efficient solution for deploying agentic AI systems in resource-constrained communication networks, balancing personalization with global knowledge transfer.