The Art of Asking: Multilingual Prompt Optimization for Synthetic Data

2025-10-22

AI Summary

Current multilingual large language models (LLMs) rely heavily on translation-based prompting, leading to English centrism, cultural misalignment, and constrained cross-lingual generalization. To address this, we propose the first systematic framework for optimizing multilingual prompt spaces—GlobalPrompt—abandoning literal translation in favor of three complementary strategies: naturalness-aware rewriting, culture-specific localization, and task-difficulty augmentation. Our method automatically optimizes prompts across 12 diverse languages using only off-the-shelf multilingual LLMs, requiring no fine-tuning or additional training. Extensive experiments demonstrate substantial improvements: +4.7% accuracy on Global-MMLU, +2.4 points on Flores XCOMET-XL (a reference-free multilingual quality metric), and a 35.3% preference win rate on mArenaHard—a rigorous multilingual benchmark assessing cultural relevance and robustness. These results confirm that our approach significantly enhances cultural alignment, adversarial robustness, and global adaptability of multilingual LLMs.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

Synthetic data has become a cornerstone for scaling large language models, yet its multilingual use remains bottlenecked by translation-based prompts. This strategy inherits English-centric framing and style and neglects cultural dimensions, ultimately constraining model generalization. We argue that the overlooked prompt space-the very inputs that define training distributions-offers a more powerful lever for improving multilingual performance. We introduce a lightweight framework for prompt-space optimization, where translated prompts are systematically transformed for Naturalness, Cultural Adaptation, and Difficulty Enhancement. Using an off-the-shelf multilingual LLM, we apply these transformations to prompts for 12 languages spanning 7 families. Under identical data conditions, our approaches achieve substantial and consistent downstream improvements over the translation-only baseline: +4.7% on Global-MMLU accuracy, +2.4% on Flores XCometXL and +35.3% wins in preferences on mArenaHard. We establish prompt-space optimization as a simple yet powerful paradigm for building multilingual LLMs that are more robust, culturally grounded, and globally capable.

Problem

• Optimizing multilingual prompts to overcome translation-based limitations
• Enhancing cultural adaptation and naturalness in synthetic data generation
• Improving multilingual LLM robustness through systematic prompt transformations

Innovation

• Optimizing prompts for naturalness, cultural adaptation, difficulty
• Lightweight framework transforming translated prompts systematically
• Improving multilingual LLM performance via prompt-space optimization