Reproduce “Large Language Models are Good Multi-lingual Learners : When LLMs Meet Cross-lingual Prompts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Teng Wang, Zhenqi He, Wing-Yin Yu, Xiaojin Fu, Xiongwei Han
Researchers from the University of Hong Kong and Huawei's Noah's Ark Lab introduced Multi-Lingual Prompt (MLPrompt), a novel strategy leveraging LLMs' multilingual capabilities to improve adherence to complex rules in structured data generation tasks. This method demonstrated an impressive 37.2% accuracy increase for GPT-4o in generating Mixed Integer Programming instances and significantly reduced error rates in text-to-SQL by strategically translating error-prone instructions into non-dominant languages.
Key insights
- LLMs, like human polyglots, can benefit from the strategic introduction of non-dominant languages to enhance focus and reasoning in dominant language contexts, particularly for error-prone instructions.
- Visualizations of attention maps confirmed that MLPrompt effectively directs the LLM's internal attention mechanism towards specific, translated, error-prone rules.
- MLPrompt is more efficient than multi-step reasoning methods, achieving improved accuracy within a single inference step (after initial prompt refinement) and outperforming them, especially for medium-scale LLMs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.