How Robust are LLMs to In-Context Majority Label Bias?

TL;DR

This paper investigates how different large language models handle majority label bias in in-context learning, revealing variability in robustness influenced by model size and prompt richness, with some models being highly resilient.

Contribution

The study provides a comprehensive analysis of the robustness of open-source LLMs to majority label bias, highlighting factors that enhance or diminish their resilience.

Findings

Robustness varies widely across models and tasks.

Certain LLMs are highly robust (~90%) to majority label bias.

Model size and prompt richness significantly impact robustness.

Abstract

In the In-Context Learning (ICL) setup, various forms of label biases can manifest. One such manifestation is majority label bias, which arises when the distribution of labeled examples in the in-context samples is skewed towards one or more specific classes making Large Language Models (LLMs) more prone to predict those labels. Such discrepancies can arise from various factors, including logistical constraints, inherent biases in data collection methods, limited access to diverse data sources, etc. which are unavoidable in a real-world industry setup. In this work, we study the robustness of in-context learning in LLMs to shifts that occur due to majority label bias within the purview of text classification tasks. Prior works have shown that in-context learning with LLMs is susceptible to such biases. In our study, we go one level deeper and show that the robustness boundary varies widely for different models and tasks, with certain LLMs being highly robust (~90%) to majority label bias. Additionally, our findings also highlight the impact of model size and the richness of instructional prompts contributing towards model robustness. We restrict our study to only publicly available open-source models to ensure transparency and reproducibility.