Enhancing In-Context Learning Performance with just SVD-Based Weight Pruning: A Theoretical Perspective

TL;DR

This paper demonstrates that SVD-based weight pruning can improve in-context learning performance in large language models, with deeper layer pruning often yielding more stable improvements, supported by theoretical analysis and experimental validation.

Contribution

It provides the first theoretical explanation for how SVD-based pruning enhances ICL, and introduces a simple, effective algorithm for improving ICL in downstream tasks.

Findings

Pruning deep layers leads to more stable ICL improvements.

SVD-based pruning enhances ICL performance on benchmark datasets.

Theoretical analysis explains the mechanisms behind pruning benefits.

Abstract

Pre-trained large language models (LLMs) based on Transformer have demonstrated striking in-context learning (ICL) abilities. With a few demonstration input-label pairs, they can predict the label for an unseen input without any parameter updates. In this paper, we show an exciting phenomenon that SVD-based weight pruning can enhance ICL performance, and more surprising, pruning weights in deep layers often results in more stable performance improvements than in shallow layers. However, the underlying mechanism of those findings still remains an open question. To reveal those findings, we conduct an in-depth theoretical analysis by presenting the implicit gradient descent (GD) trajectories of ICL and giving the mutual information based generalization bounds of ICL via full implicit GD trajectories. This helps us reasonably explain the surprising experimental findings. Besides, based on all our experimental and theoretical insights, we intuitively propose a simple, model-compression and derivative-free algorithm for downstream tasks in enhancing ICL inference. Experiments on benchmark datasets and open source LLMs display the method effectiveness\footnote{The code is available at \url{https://github.com/chen123CtrlS/EnhancingICL_SVDPruning}.}.