Do Influence Functions Work on Large Language Models?

TL;DR

This paper systematically evaluates influence functions on large language models and finds they perform poorly due to approximation errors, convergence issues, and fundamental definition problems, indicating the need for alternative methods.

Contribution

It provides the first comprehensive assessment of influence functions on LLMs, revealing their limitations and highlighting the necessity for new approaches.

Findings

Influence functions perform poorly on LLMs across multiple tasks.

Approximation errors and convergence issues hinder influence function effectiveness.

Changes in model parameters do not reliably indicate changes in LLM behavior.

Abstract

Influence functions are important for quantifying the impact of individual training data points on a model's predictions. Although extensive research has been conducted on influence functions in traditional machine learning models, their application to large language models (LLMs) has been limited. In this work, we conduct a systematic study to address a key question: do influence functions work on LLMs? Specifically, we evaluate influence functions across multiple tasks and find that they consistently perform poorly in most settings. Our further investigation reveals that their poor performance can be attributed to: (1) inevitable approximation errors when estimating the iHVP component due to the scale of LLMs, (2) uncertain convergence during fine-tuning, and, more fundamentally, (3) the definition itself, as changes in model parameters do not necessarily correlate with changes in LLM behavior. Thus, our study suggests the need for alternative approaches for identifying influential samples.