Efficiently Computing Susceptibility to Context in Language Models

TL;DR

This paper introduces Fisher susceptibility, a fast and efficient method to measure how sensitive language models are to context changes, enabling better analysis of model behavior.

Contribution

We propose Fisher susceptibility, an efficient alternative to Monte Carlo methods for quantifying language model sensitivity to context, validated across diverse domains.

Findings

Fisher susceptibility is 70 times faster than Monte Carlo approximation.

Larger models are as susceptible as smaller ones.

Fisher susceptibility closely matches Monte Carlo estimates.

Abstract

One strength of modern language models is their ability to incorporate information from a user-input context when answering queries. However, they are not equally sensitive to the subtle changes to that context. To quantify this, Du et al. (2024) gives an information-theoretic metric to measure such sensitivity. Their metric, susceptibility, is defined as the degree to which contexts can influence a model's response to a query at a distributional level. However, exactly computing susceptibility is difficult and, thus, Du et al. (2024) falls back on a Monte Carlo approximation. Due to the large number of samples required, the Monte Carlo approximation is inefficient in practice. As a faster alternative, we propose Fisher susceptibility, an efficient method to estimate the susceptibility based on Fisher information. Empirically, we validate that Fisher susceptibility is comparable to Monte Carlo estimated susceptibility across a diverse set of query domains despite its being $70\times$ faster. Exploiting the improved efficiency, we apply Fisher susceptibility to analyze factors affecting the susceptibility of language models. We observe that larger models are as susceptible as smaller ones.