Towards Informative Few-Shot Prompt with Maximum Information Gain for In-Context Learning

TL;DR

This paper proposes a method to select the most informative examples for in-context learning in large language models by maximizing information gain, leading to more stable and effective few-shot prompts.

Contribution

It introduces a novel approach to quantify and maximize information gain in example selection, and addresses template bias with a calibration strategy.

Findings

Achieves 14.3% average relative improvement across six classification tasks

Reduces variance in in-context learning performance

Enhances stability and fairness in few-shot prompting

Abstract

Large Language models (LLMs) possess the capability to engage In-context Learning (ICL) by leveraging a few demonstrations pertaining to a new downstream task as conditions. However, this particular learning paradigm suffers from high instability stemming from substantial variances induced by factors such as the input distribution of selected examples, their ordering, and prompt formats. In this work, we demonstrate that even when all these factors are held constant, the random selection of examples still results in high variance. Consequently, we aim to explore the informative ability of data examples by quantifying the Information Gain (IG) obtained in prediction after observing a given example candidate. Then we propose to sample those with maximum IG. Additionally, we identify the presence of template bias, which can lead to unfair evaluations of IG during the sampling process. To mitigate this bias, we introduce Calibration Before Sampling strategy. The experimental results illustrate that our proposed method can yield an average relative improvement of 14.3% across six classification tasks using three LLMs.