In-Person Poster presentation / poster accept
Language models are multilingual chain-of-thought reasoners
Freda Shi · Mirac Suzgun · Markus Freitag · Xuezhi Wang · Suraj Srivats · Soroush Vosoughi · Hyung Won Chung · Yi Tay · Sebastian Ruder · Denny Zhou · Dipanjan Das · Jason Wei
MH1-2-3-4 #37
Keywords: [ Applications ] [ large language model ] [ reasoning ] [ multilingual ]
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 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 AnonymousLink and the supplementary material.