Are LLM-Judges Robust to Expressions of Uncertainty? Investigating the effect of Epistemic Markers on LLM-based Evaluation

TL;DR

This paper introduces EMBER, a benchmark to evaluate how robust LLM-based judges are when assessing outputs containing epistemic markers, revealing a bias against uncertainty expressions and highlighting a robustness issue.

Contribution

The study presents EMBER, the first benchmark for testing LLM-judges' robustness to epistemic markers, and demonstrates their vulnerability to bias caused by such markers.

Findings

LLM-judges show bias against epistemic markers.

All tested LLM-judges are affected by epistemic markers.

Bias is stronger against markers expressing uncertainty.

Abstract

In line with the principle of honesty, there has been a growing effort to train large language models (LLMs) to generate outputs containing epistemic markers. However, evaluation in the presence of epistemic markers has been largely overlooked, raising a critical question: Could the use of epistemic markers in LLM-generated outputs lead to unintended negative consequences? To address this, we present EMBER, a benchmark designed to assess the robustness of LLM-judges to epistemic markers in both single and pairwise evaluation settings. Our findings, based on evaluations using EMBER, reveal that all tested LLM-judges, including GPT-4o, show a notable lack of robustness in the presence of epistemic markers. Specifically, we observe a negative bias toward epistemic markers, with a stronger bias against markers expressing uncertainty. This suggests that LLM-judges are influenced by the presence of these markers and do not focus solely on the correctness of the content.