Model Merging via Data-Free Covariance Estimation

TL;DR

This paper introduces a data-free covariance estimation method for model merging that minimizes task interference, reducing data dependency and computational costs, and demonstrating superior performance on vision and language benchmarks.

Contribution

It proposes a novel data-free covariance estimation technique for model merging, improving efficiency and effectiveness over previous data-free methods.

Findings

Outperforms previous data-free merging methods on vision and language benchmarks.

Estimates covariance matrices directly from difference matrices, eliminating data dependency.

Reduces computational costs compared to data-dependent approaches.

Abstract

Model merging provides a way of cheaply combining individual models to produce a model that inherits each individual's capabilities. While some merging methods can approach the performance of multitask training, they are often heuristically motivated and lack theoretical justification. A principled alternative is to pose model merging as a layer-wise optimization problem that directly minimizes interference between tasks. However, this formulation requires estimating per-layer covariance matrices from data, which may not be available when performing merging. In contrast, many of the heuristically-motivated methods do not require auxiliary data, making them practically advantageous. In this work, we revisit the interference minimization framework and show that, under certain conditions, covariance matrices can be estimated directly from difference matrices, eliminating the need for data while also reducing computational costs. We validate our approach across vision and language benchmarks on models ranging from 86M parameters to 7B parameters, outperforming previous data-free state-of-the-art merging methods