What Did I Do Wrong? Quantifying LLMs' Sensitivity and Consistency to Prompt Engineering

TL;DR

This paper introduces metrics to quantify how sensitive and consistent large language models are to prompt variations, aiding developers in debugging and improving prompt engineering for more reliable AI outputs.

Contribution

It proposes novel sensitivity and consistency metrics for classification tasks, providing tools to analyze and improve LLM robustness beyond traditional performance measures.

Findings

Sensitivity varies significantly with prompt rephrasing.

Consistency correlates with model robustness.

Metrics help identify failure modes in LLM predictions.

Abstract

Large Language Models (LLMs) changed the way we design and interact with software systems. Their ability to process and extract information from text has drastically improved productivity in a number of routine tasks. Developers that want to include these models in their software stack, however, face a dreadful challenge: debugging LLMs' inconsistent behavior across minor variations of the prompt. We therefore introduce two metrics for classification tasks, namely sensitivity and consistency, which are complementary to task performance. First, sensitivity measures changes of predictions across rephrasings of the prompt, and does not require access to ground truth labels. Instead, consistency measures how predictions vary across rephrasings for elements of the same class. We perform an empirical comparison of these metrics on text classification tasks, using them as guideline for understanding failure modes of the LLM. Our hope is that sensitivity and consistency will be helpful to guide prompt engineering and obtain LLMs that balance robustness with performance.