Large Language Models are Good Multi-lingual Learners : When LLMs Meet Cross-lingual Prompts

TL;DR

The paper tackles generating realistic MIP instances for industrial optimization using LLMs, addressing rule-omission errors in long natural-language prompts. It introduces MLPrompt, a multilingual prompting strategy that translates error-prone rules into non-dominant languages to steer LLM attention and improve JSON-structured data generation. Through experiments on ComplexOR and a Text-to-SQL extension, MLPrompt outperforms CoT, ToT, and SC across small, medium, and large LLMs, with notable gains from Mandarin translations and faster generation. The work demonstrates a practical bridge to autonomous MIP instance generation in industrial pipelines and suggests broader applicability of cross-lingual prompting to structured data synthesis.

Abstract

With the advent of Large Language Models (LLMs), generating rule-based data for real-world applications has become more accessible. Due to the inherent ambiguity of natural language and the complexity of rule sets, especially in long contexts, LLMs often struggle to follow all specified rules, frequently omitting at least one. To enhance the reasoning and understanding of LLMs on long and complex contexts, we propose a novel prompting strategy Multi-Lingual Prompt, namely MLPrompt, which automatically translates the error-prone rule that an LLM struggles to follow into another language, thus drawing greater attention to it. Experimental results on public datasets across various tasks have shown MLPrompt can outperform state-of-the-art prompting methods such as Chain of Thought, Tree of Thought, and Self-Consistency. Additionally, we introduce a framework integrating MLPrompt with an auto-checking mechanism for structured data generation, with a specific case study in text-to-MIP instances. Further, we extend the proposed framework for text-to-SQL to demonstrate its generation ability towards structured data synthesis.

Figures (7)

• Figure 1: Semantic illustrations of the potential of LLMs in generating data following real distributions. (a) Distribution of Constraint Coefficients in Real-World Factory Location Problems cornuejols1977exceptional. (b) Distribution of Constraint Coefficients generated by random simulation. (c) Distribution of Constraint Coefficients generated by GPT-$4$achiam2023gpt with different hyper-parameters. For each case: (demand points, candidate locations) = (4, 4), (5, 4).
• Figure 2: Illustration of the MIP instance generation process using key modeling components, including sets, parameters, variables, constraints, and objective functions. Only when compliant with the parser rules can the parser generate the modeling code, and the modeling tools produce instances.
• Figure 3: The dominant language refers to the majority language in the pretraining dataset, while "others" refer to all remaining languages. This figure demonstrates the language imbalance phenomenon in the pretraining data of GPT-$3.5$brown2020language, Llama-$3$llama3, Deepseek-V$2$deepseek, Bloom bloom, Nemotron-4 adler2024nemotron, OPT zhang2022opt, and Falcon almazrouei2023falcon.
• Figure 4: The workflow of the proposed framework for structured data generation. The process includes prompt construction with predefined rules, data generation by the LLM, evaluation of compliance with the rules, and iterative refinement of the prompt by translating rules into other languages if necessary.
• Figure 5: A demo of how MLPrompt builds prompts.