Evaluating and Mitigating LLM-as-a-judge Bias in Communication Systems

2025-10-14

AI Summary

This work investigates bias in large language models (LLMs) when deployed as automated content evaluators in communication systems, exposing how implicit and explicit biases in their scoring processes undermine fairness and user trust. Methodologically, we systematically identify 11 categories of judgment bias and conduct empirical analysis using GPT-Judge and JudgeLM across four dimensions: pointwise scoring, fine-grained criterion-level analysis, bias-aware response fine-tuning, and cross-dataset evaluation. Key findings include: (i) mainstream LLM judges exhibit robustness to biased inputs; (ii) refining scoring criteria significantly improves fairness; (iii) training data bias severely degrades evaluation reliability; and (iv) task difficulty strongly correlates with scoring outcomes. Based on these insights, we propose four transferable bias-mitigation strategies—criterion decomposition, bias-aware calibration, data balancing, and difficulty-aware normalization—to advance trustworthy, equitable AI judging paradigms.

Technology Category

• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Philosophy and Ethics of AI: Bias, Fairness & Equity

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Social Networks and Social Media: Fairness and bias in social network and social media analysis

Abstract

Large Language Models (LLMs) are increasingly being used to autonomously evaluate the quality of content in communication systems, e.g., to assess responses in telecom customer support chatbots. However, the impartiality of these AI "judges" is not guaranteed, and any biases in their evaluation criteria could skew outcomes and undermine user trust. In this paper, we systematically investigate judgment biases in two LLM-as-a-judge models (i.e., GPT-Judge and JudgeLM) under the point-wise scoring setting, encompassing 11 types of biases that cover both implicit and explicit forms. We observed that state-of-the-art LLM judges demonstrate robustness to biased inputs, generally assigning them lower scores than the corresponding clean samples. Providing a detailed scoring rubric further enhances this robustness. We further found that fine-tuning an LLM on high-scoring yet biased responses can significantly degrade its performance, highlighting the risk of training on biased data. We also discovered that the judged scores correlate with task difficulty: a challenging dataset like GPQA yields lower average scores, whereas an open-ended reasoning dataset (e.g., JudgeLM-val) sees higher average scores. Finally, we proposed four potential mitigation strategies to ensure fair and reliable AI judging in practical communication scenarios.

Problem

• Evaluating bias in LLM judges used for content quality assessment
• Investigating 11 types of implicit and explicit judgment biases
• Developing mitigation strategies for fair AI judging in communications

Innovation

• Systematically investigated 11 types of LLM judgment biases
• Proposed four mitigation strategies for fair AI judging
• Found scoring rubrics enhance robustness to biased inputs