Multilingual Prompt Localization for Agent-as-a-Judge: Language and Backbone Sensitivity in Requirement-Level Evaluation
Alhasan Mahmood, Samir Abdaljalil, Hasan Kurban
Why It Matters
What makes this one worth your time
Understanding language sensitivity in AI evaluations can improve the design of benchmarks and enhance the reliability of AI systems in multilingual contexts.
Language choice significantly influences AI judge performance across diverse languages.
Summary
The paper investigates how the language of evaluation affects the performance of AI judges across multiple languages and frameworks, revealing significant interactions between language and model backbone in requirement-level evaluations.
Key contributions
- Localization of the Agent-as-a-Judge prompt stack to five languages.
- Empirical evaluation of 55 DevAI tasks across multiple frameworks and judge backbones.
- Release of full requirement-level judgments and runtime statistics for reproducibility.
Notable insights
- Localizing judge-side instructions is crucial for maintaining evaluation quality across languages.
- No single AI backbone consistently outperforms others across different languages, indicating the need for tailored approaches.
Possible limitations
- Potential biases in language representation or task selection are not addressed.
- The abstract does not specify the criteria for selecting the five languages or the implications of modest inter-backbone agreement.
Abstract
arXiv:2604.04532v2 Announce Type: replace-cross Abstract: Evaluation language is typically treated as a fixed English default in agentic code benchmarks, yet we show that changing the judge's language can invert backbone rankings. We localize the Agent-as-a-Judge prompt stack to five typologically diverse languages (English, Arabic, Turkish, Chinese, Hindi) and evaluate 55 DevAI development tasks across three developer-agent frameworks and six judge backbones, totaling 4950 judge runs. The central finding is that backbone and language interact: GPT-4o achieves the highest satisfaction in English (44.72\%), while Gemini leads in Arabic (51.72\%, $p<0.001$ vs.\ GPT-4o) and Hindi (53.22\%). No single backbone dominates across all languages, and inter-backbone agreement on individual requirement judgments is modest (Fleiss' $\kappa \leq 0.231$). A controlled ablation further shows that localizing judge-side instructions, not just benchmark content, can be decisive: Hindi satisfaction drops from 42.8\% to 23.2\% under partial localization. These results indicate that language should be treated as an explicit evaluation variable in agentic benchmarks. Full requirement-level judgments and runtime statistics are released for reproducibility.