This paper explores the capacity of large language models (LLMs) for reasoning in math word problems (MWPs) by employing symbolic representations of numeric problems. The authors find that the GPT-3’s davinci-002 model demonstrates strong zero-shot accuracy on symbolic MWPs. They introduce a self-prompting approach to align symbolic reasoning with the numeric answer, thereby making the LLMs more interpretable. Interestingly, self-prompting also enhances symbolic accuracy, revealing an ensembling effect. The research is supported by a symbolic version of the SVAMP dataset called SVAMP-Sym, created for future research on symbolic math problems.
Publication date: August 3, 2023
Project Page: Not available
Paper: https://arxiv.org/pdf/2308.01906.pdf
