Vision–Language Models (VLMs) have recently attracted considerable attention in Federated Learning (FL) due to their strong and robust performance. In particular, few-shot adaptation with pre-trained VLMs like CLIP enhances the performance of downstream tasks. However, existing methods still suffer from substantial communication overhead, high local computational demands, and suboptimal performance under non-IID user data. To simultaneously address all those limitations, we propose NormFit, a lightweight solution that selectively fine-tunes only a very small portion of the model parameters, specifically only the Pre-LayerNorm parameters of the vision encoder within a VLM. Overcoming the existing tradeoff between performance and communication/computation efficiency in few-shot FL, NormFit sets a new benchmark by simultaneously achieving superior accuracy and substantially reduced communication and computational demands. Theoretically, we show that NormFit yields a considerably smaller generalization gap compared to tuning all LayerNorm parameters. Importantly, NormFit can function effectively as a standalone solution or integrate seamlessly with existing few-shot fine-tuning methods to further enhance their performance. Notably, NormFit offers implementation simplicity, achieving these improvements without any algorithmic modifications, changes to the underlying model architecture, or the addition of external parameters.

One of the most critical challenges of federated learning (FL) is to send data efficiently and reliably over the noisy wireless channels between the clients and server to achieve target learning accuracy as fast as possible. To achieve this goal, we design effective error correction coded FL with managed retransmissions. Rather than using Shannon capacity as the performance measure to design the communication mechanisms for FL, our approach relies critically on learning accuracy. Our fundamental idea is based on the observation that Stochastic Gradient Decent (SGD) and its family can tolerate some errors in the course of training. Inspired by this, to reduce the communication burden without degrading the learning accuracy, our FL framework with Managed Redundancy (FL-MR) has two phases: (i) the No-Retransmission phase, where retransmissions are never performed even in case of erroneous decoding of data and (ii) the Select Retransmission phase, where only some carefully selected data packets are retransmitted. Our extensive simulation results demonstrate that the proposed coded FL system achieves target accuracies much faster than the baseline coded approach.