Fairness or Fluency? An Investigation into Language Bias of Pairwise LLM-as-a-Judge

TL;DR

The paper investigates language bias in pairwise LLM-as-a-judge under two setups: same-language and inter-language judging, using MMMLU across 14 languages and a mix of models. It analyzes whether bias correlates with perplexity differences via $R^2$ decomposition and nested $F$-tests, and finds substantial disparities across languages, with European languages performing best in same-language tasks and a strong English preference in inter-language tasks driven largely by the answer language. The study demonstrates that language identity explains additional variance beyond perplexity, especially for low-resource languages, indicating language bias is not merely a reflection of model fluency. These results have practical implications for multilingual evaluation and the design of fair, robust LLM-based judging systems, highlighting the need to address language-specific biases beyond perplexity considerations.

Abstract

Recent advances in Large Language Models (LLMs) have incentivized the development of LLM-as-a-judge, an application of LLMs where they are used as judges to decide the quality of a certain piece of text given a certain context. However, previous studies have demonstrated that LLM-as-a-judge can be biased towards different aspects of the judged texts, which often do not align with human preference. One of the identified biases is language bias, which indicates that the decision of LLM-as-a-judge can differ based on the language of the judged texts. In this paper, we study two types of language bias in pairwise LLM-as-a-judge: (1) performance disparity between languages when the judge is prompted to compare options from the same language, and (2) bias towards options written in major languages when the judge is prompted to compare options of two different languages. We find that for same-language judging, there exist significant performance disparities across language families, with European languages consistently outperforming African languages, and this bias is more pronounced in culturally-related subjects. For inter-language judging, we observe that most models favor English answers, and that this preference is influenced more by answer language than question language. Finally, we investigate whether language bias is in fact caused by low-perplexity bias, a previously identified bias of LLM-as-a-judge, and we find that while perplexity is slightly correlated with language bias, language bias cannot be fully explained by perplexity only.

Figures (21)

• Figure 1: Illustration of language bias in the same-language judging and inter-language judging scenario.
• Figure 2: Same-language judging performance of tested models on MMMLU. Models are sorted by average performance across all languages (top to bottom), languages are sorted by average performance across all models (left to right).
• Figure 3: Question Effect versus Answer Effect in same-language judging. Points above the diagonal indicate greater sensitivity to answer language; points below indicate greater sensitivity to question language.
• Figure 4: Accuracy gap between English and target language by subject category for Prometheus2-8x7B. Bars represent performance difference across STEM, Humanities, Social Sciences, and Other subjects for each language.
• Figure 5: Inter-language judging performance across models and languages. Each cell shows accuracy when comparing English and target language answers, averaged across all four configurations. Models are sorted by average performance across all languages (top to bottom), languages are sorted by average performance across all models (left to right).