A way to personalize federated adapter tuning that treats each client's whole sequence of past adapter updates as signal instead of using only the latest round. The server keeps a hypernetwork called a sequential learner that consumes the stacked per-round, per-client update deltas and emits a per-client calibration, which is added to the FedAvg-aggregated global adapter to form that client's personalized adapter for the next round. Because the adapter's parameter dimension is treated as a batch dimension, the learner's size depends on the client count and sequence length rather than on adapter size, and it is trained at the server with no access to client data by using each received update delta as a proxy for the gradient of that client's local loss. A selective state space model is used as the learner so that cross-client interactions enter the input-dependent recurrence parameters while cross-step dependencies accumulate in the hidden state; a capped maximum sequence length and a warm-up period before personalization begins keep the scheme bounded and stable. The concept teaches why using update history is analogous to adding momentum to gradient descent, and how to isolate the contribution of global aggregation, cross-step modelling and cross-client modelling by ablation.
Personalized federated learning (PFL) gives each client its own model instead of one global model, because non-IID and unbalanced client data make a single shared model perform poorly locally. When t…