On the Advance of Making Language Models Better Reasoners

Proposes DIVERSE, which, combined with davinci-002, achieves new SotA on various reasoning benchmarks (e.g., GSM8K 74.4% -> 83.2%), outperforming the 540B PaLM.

This paper highlights the potential of language models for improving reasoning tasks and proposes a new approach, DiVeRSe, which achieved state-of-the-art performance on multiple reasoning benchmarks. Implementing language models like DiVeRSe in business processes can increase operational efficiency and accuracy.