A framework that merges independently trained neural networks with different architectures into one multi-task model without any retraining, handling both depth heterogeneity (different numbers of layers) and width heterogeneity (different layer sizes). Depth is reconciled by a layer-alignment strategy that segments a deeper model into groups of consecutive layers with similar representations so they can be matched to a shallower model's layers; width is reconciled by an elastic neuron-zipping algorithm that projects weights of differing widths onto a shared dimensional space. It extends training-free model merging beyond the usual homogeneous-architecture assumption while matching homogeneous-merge performance on vision and NLP tasks.
Model merging has attracted significant attention as a powerful paradigm for model reuse, facilitating the integration of task-specific models into a singular, versatile framework endowed with multif…