Task Singular Vectors: Reducing Task Interference in Model Merging

TL;DR

This paper introduces Task Singular Vectors (TSV), a layer-level approach to model merging that reduces task interference and compresses task matrices, leading to improved merging performance without additional training.

Contribution

It proposes TSV-Compress for low-rank compression of task matrices and TSV-Merge for interference-aware model merging, advancing model merging techniques.

Findings

TSV-Compress retains 99% accuracy at 10% original size.

TSV-Merge outperforms existing model merging methods.

Layer task matrices are often low-rank, enabling effective compression.

Abstract

Task Arithmetic has emerged as a simple yet effective method to merge models without additional training. However, by treating entire networks as flat parameter vectors, it overlooks key structural information and is susceptible to task interference. In this paper, we study task vectors at the layer level, focusing on task layer matrices and their singular value decomposition. In particular, we concentrate on the resulting singular vectors, which we refer to as Task Singular Vectors (TSV). Recognizing that layer task matrices are often low-rank, we propose TSV-Compress (TSV-C), a simple procedure that compresses them to 10% of their original size while retaining 99% of accuracy. We further leverage this low-rank space to define a new measure of task interference based on the interaction of singular vectors from different tasks. Building on these findings, we introduce TSV-Merge (TSV-M), a novel model merging approach that combines compression with interference reduction, significantly outperforming existing methods.