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

TL;DR

EMBER introduces a benchmark to quantify how LLM-judges evaluate outputs in the presence of epistemic markers, addressing a gap in honesty-alignment research. It defines EMBERQA and EMBERIF to probe single and pairwise evaluations, with outputs augmented by strengtheners and weakeners drawn from real-world usage. Across five LLM-judges, results show a consistent bias against epistemic markers, with stronger effects for weakeners and a robustness trend that improves with model scale but remains incomplete, as measured by the $VSR$ (Verdict Switch Rate). Human judges show little to no bias toward markers, highlighting a gap between human and machine evaluation and underscoring the need for robust, fair evaluation protocols. The findings have real-life implications, including potential misranking of stronger models when their outputs include uncertainty expressions.

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.

Figures (5)

• Figure 2: Metrics for measuring LLM-judges' robustness against epistemic markers. Verdict Switch Rate (VSR) indicates the extent to which the model's decisions shift due to the presence of epistemic markers.
• Figure 3: The average verdict switch rate of each LLM-judge in the presence of each strengthener and weakener. (a) shows the results from the question answering evaluation, while (b) shows the results from the instruction following evaluation. A lower value indicates greater robustness.
• Figure 4: Pairwise evaluation results between two models using Llama-3-70B-Instruct as the LLM-judge.
• Figure 5: Interface of human-judge evaluation.
• Figure 6: Instruction given to the human annotators.