Judge Language and Backbone Interaction Affects Agent Evaluation

Localizing Agent-as-a-Judge prompts to non-English languages like Arabic and Hindi can invert backbone model rankings. While GPT-4o leads in English, Gemini outperforms it in Arabic and Hindi satisfaction scores, demonstrating that judge-side localization is critical for accurate multilingual evaluation.