FedLAFP splits federated fine-tuning into two jobs instead of forcing one adapter to do both.
Researchers propose FedLAFP, a framework for federated parameter-efficient fine-tuning, the technique that lets phones, hospitals, or other clients adapt a shared pretrained model without uploading raw data. Instead of using one low-rank adapter for both shared knowledge and client-specific quirks, FedLAFP pairs two: a compact LoRA branch that gets averaged across all clients, and a private RandLoRA branch, built from fixed random low-rank bases plus learned scaling coefficients, that never leaves the device. A per-client, per-layer mixing coefficient blends the two, and only the shared LoRA weights are ever transmitted. Across four visual recognition benchmarks, the approach reached an average personalized accuracy of 86.93%, beating the best federated LoRA baseline by 1.30 percentage points.
Federated fine-tuning has mostly reused the same low-rank recipe for both the parts that need to generalize and the parts that need to stay idiosyncratic, and that mismatch is a likely reason personalization has lagged. FedLAFP's contribution is not a bigger model or more data. It is recognizing that aggregation and personalization want different math, then building the plumbing so only the aggregation-friendly piece travels over the network.
A 1.3-point bump on four benchmarks is solid, not spectacular, and "full-rank-capable" here means a richer private adapter, not an actually full-rank model. The real test is whether this holds up past image classification, in noisier, real-world federated deployments.