Language Models are Multilingual Chain-of-Thought Reasoners

TL;DR

This paper demonstrates that large language models exhibit emerging multilingual reasoning abilities, including chain-of-thought reasoning, across diverse languages and tasks, with performance improving as models scale.

Contribution

Introduction of the MGSM benchmark for multilingual reasoning and evidence that large models develop strong reasoning skills across multiple languages.

Findings

Models improve in multilingual reasoning with scale.

Strong reasoning abilities in underrepresented languages.

Multilingual reasoning extends to various tasks.

Abstract

We evaluate the reasoning abilities of large language models in multilingual settings. We introduce the Multilingual Grade School Math (MGSM) benchmark, by manually translating 250 grade-school math problems from the GSM8K dataset (Cobbe et al., 2021) into ten typologically diverse languages. We find that the ability to solve MGSM problems via chain-of-thought prompting emerges with increasing model scale, and that models have strikingly strong multilingual reasoning abilities, even in underrepresented languages such as Bengali and Swahili. Finally, we show that the multilingual reasoning abilities of language models extend to other tasks such as commonsense reasoning and word-in-context semantic judgment. The MGSM benchmark is publicly available at https://github.com/google-research/url-nlp.