Reproduce “Model Merging: Foundations and Algorithms” on OpenResearch

About this paper

Donato Crisostomi

This doctoral thesis establishes theoretical and algorithmic foundations for model merging, a paradigm for combining independently trained neural networks into a single unified model without additional training data. It presents methods that achieve significant accuracy gains in both single- and multi-task scenarios, reduce computational costs for model composition, and enable new capabilities like input-adaptive merging and cross-lingual knowledge transfer.

Key insights

• Neuron permutation symmetries are a primary barrier to directly merging single-task models, and cycle-consistent alignment provides a robust geometric framework to address this.
• Task vectors, which represent the parameter difference between fine-tuned and base models, can be formally linked to approximate scaled negative gradients, providing a theoretical explanation for the effectiveness of task arithmetic.
• Per-layer task matrices exhibit an intrinsic low-rank structure, meaning learned information can be efficiently represented by a small fraction of singular components, which can be decorrelated to minimize task interference.

Summary by alphaXiv.