Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck

Researchers introduce DIBJudge, a new framework to address systematic bias in large language models that favor machine-translated text over human-authored content in multilingual evaluations. The solution uses variational information compression to isolate bias factors and improve LLM judgment accuracy across languages.

• →Large language models exhibit translationese bias, systematically favoring machine-translated text over human-authored references in multilingual evaluations.
• →The bias is particularly pronounced in low-resource languages and stems from spurious correlations with English alignment and cross-lingual predictability.
• →DIBJudge framework uses variational information compression to learn judgment-critical representations while isolating bias factors.
• →The approach incorporates cross-covariance penalty to suppress statistical dependence between robust and bias representations.
• →Extensive evaluations show DIBJudge consistently outperforms existing baselines and substantially reduces translationese bias.