Reproduce “Cross-lingual Prompting: Improving Zero-shot Chain-of-Thought Reasoning across Languages” on OpenResearch

About this paper

Libo Qin, Qiguang Chen, Fuxuan Wei, Shijue Huang, Wanxiang Che

This research introduces Cross-lingual Prompting (CLP) and Cross-lingual Self-consistent Prompting (CLSP) to enable zero-shot Chain-of-Thought reasoning in Large Language Models across multiple languages. The methods consistently improve reasoning accuracy on multilingual benchmarks like MGSM, achieving up to 76.7% accuracy with CLSP using GPT-3.5, and demonstrating enhanced reasoning quality and broad applicability across various LLMs and tasks.

Key insights

• The two-stage interactive prompting approach of CLP significantly outperforms single-turn prompting, indicating that explicit dialogue history better elicits LLM capabilities for cross-lingual tasks.
• CLP is not merely a translation method; it employs diverse strategies like problem restatement, key information extraction, and preliminary solution steps, leading to higher quality and more informative reasoning paths.
• CLSP's performance gains are linked to its ability to generate and integrate diverse reasoning knowledge from multiple languages, with the effectiveness influenced by the resource levels (pretraining data quantity) of the integrated languages.

Summary by alphaXiv.