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

2024-10-28

arXiv.org

AI Summary

This work investigates the robustness of large language models (LLMs) acting as evaluators (“LLM-judges”) to cognitive markers—such as “possibly” or “uncertain”—in generated text. We identify a systematic negative bias: LLM-judges consistently underestimate the quality of outputs containing uncertainty expressions. To address this, we introduce ERMER, the first benchmark explicitly designed to probe cognitive marker sensitivity. ERMER combines prompt engineering with controllable perturbations to generate cognitively marked variants and employs both single-turn scoring and pairwise comparison evaluation paradigms across mainstream models (e.g., GPT-4o). Experiments reveal statistically significant negative bias across all tested LLM-judges (mean score deviation: 12.3%, *p* < 0.001), indicating that evaluations are driven more by linguistic surface cues than semantic correctness. This study provides the first empirical evidence of cognitive marker sensitivity in LLM-based evaluation, offering critical insights and methodological foundations for developing trustworthy AI assessment frameworks.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

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.

Problem

• Assessing LLM-judges' robustness to epistemic markers
• Investigating bias in LLM evaluations with uncertainty expressions
• Evaluating impact of epistemic markers on LLM-generated outputs

Innovation

• EMBER benchmark tests LLM-judges robustness
• LLM-judges show bias against uncertainty markers
• Evaluates single and pairwise epistemic marker effects