15 Oct 2025
Weight Weaving introduces a parameter pooling framework that enables data-free deep neural network model merging by marginalizing over a user-defined search space of scaling factors. This approach consistently enhances the performance of existing state-of-the-art merging methods, achieving average accuracy gains up to 15.9 percentage points in continual learning and domain generalization tasks.
There are no more papers matching your filters at the moment.
