On the Relation between Sensitivity and Accuracy in In-context Learning

TL;DR

This paper investigates the sensitivity of in-context learning models to prompt perturbations, revealing that sensitivity correlates negatively with accuracy, and introduces a selective prediction method that improves reliability by abstaining from sensitive predictions.

Contribution

It uncovers the impact of label bias on ICL sensitivity, demonstrates the negative correlation between sensitivity and accuracy, and proposes extsc{SenSel}, a method for improved prediction reliability.

Findings

Sensitivity is underestimated due to label bias.

Higher sensitivity correlates with lower accuracy.

extsc{SenSel} outperforms baseline abstention methods.

Abstract

In-context learning (ICL) suffers from oversensitivity to the prompt, making it unreliable in real-world scenarios. We study the sensitivity of ICL with respect to multiple perturbation types. First, we find that label bias obscures the true sensitivity, and therefore prior work may have significantly underestimated ICL sensitivity. Second, we observe a strong negative correlation between ICL sensitivity and accuracy: predictions sensitive to perturbations are less likely to be correct. Motivated by these findings, we propose \textsc{SenSel}, a few-shot selective prediction method that abstains from sensitive predictions. Experiments on ten classification datasets show that \textsc{SenSel} consistently outperforms two commonly used confidence-based and entropy-based baselines on abstention decisions.