LLM Evaluators are Biased across Languages

2026-07-15

AI Summary

This study addresses the often-assumed language neutrality of multilingual large language model (LLM) evaluators, revealing significant linguistic bias that systematically inflates scores for low-resource languages, thereby compromising fairness in downstream tasks such as safety filtering. The authors construct the first semantic-aligned instruction–response dataset spanning 23 languages and conduct a systematic evaluation of eight LLM evaluators across diverse architectures and training paradigms. Their analysis demonstrates a strong correlation between evaluation scores and language resource availability, with up to a 43% disparity in acceptance rates for low-resource languages. This bias persists beyond explanations based on content difficulty or model uncertainty, and high pairwise accuracy masks substantial cross-lingual inconsistency. The work uncovers structural biases in multilingual evaluation and provides an empirical foundation for fairer assessment through multilingual datasets, uncertainty metrics, and mixed-language analyses.

Abstract

LLM evaluators (trained reward models and prompted LLM-as-a-Judge) are routinely validated via pairwise accuracy. In a multilingual setting, this operates under the premise that high pairwise accuracy implies reliable, language-neutral scoring. We show that this assumption does not hold. We conduct experiments with semantically identical instruction-response pairs across 23 languages, and find that multilingual evaluators assign significantly different scores to different evaluation languages. The bias is statistically significant and consistent across eight open-weight evaluators of different architectures and training paradigms, persists in frontier judges, and is strongly correlated with language resource level: lower-resource languages are scored more generously. Meanwhile, these biases are invisible to pairwise accuracy: evaluators achieve above 90% pairwise accuracy, yet have up to 43% difference in acceptance rate across languages under a global decision threshold, meaning, for instance, that harmful content in lower-resource languages is more likely to pass safety filters. Per-language thresholds would require language identification, which can be defeated by code-switched prompts. We then investigate why lower-resource languages receive higher rather than lower scores, and we find that model uncertainty is linked with the effect: models tend to give higher scores when less confident, both under negative log-likelihood and under token-free uncertainty measures; however, language identity remains a significant predictor after controlling for uncertainty, and the bias cannot be explained away by content difficulty alone, but is a structural, language-level misalignment.

Problem

• LLM evaluators
• multilingual bias
• language resource disparity
• safety evaluation
• scoring inconsistency

Innovation

• language bias
• LLM-as-a-Judge
• multilingual evaluation
• model uncertainty
• safety filtering