Lower-Resource, Higher Scores: Language Bias in LLM Evaluators
Ej Zhou, Lucas Resck, Zheng Hui, Anna Korhonen
Why It Matters
What makes this one worth your time
Understanding and addressing language bias in LLM evaluators is crucial for ensuring fair and reliable AI systems across diverse linguistic contexts.
The study reveals significant language bias in LLM evaluators, affecting multilingual fairness.
Summary
The paper investigates the bias in language model evaluators across different languages, showing that these evaluators assign different scores to semantically identical content in different languages, with lower-resource languages receiving more generous scores. This bias is not detected by pairwise accuracy and is linked to model uncertainty.
Key contributions
- Demonstrated significant language bias in LLM evaluators across 23 languages.
- Linked model uncertainty to higher scores in lower-resource languages.
Notable insights
- Lower-resource languages receive higher scores due to model uncertainty.
- Language identity remains a significant predictor of bias even after controlling for uncertainty.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2607.14480v2 Announce Type: replace 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.