.png){kind=icon}
Return
CRMR: A Collaborative Multistep Reasoning Framework for Solving Mathematical Problems
.png){kind=icon}
PRE
AI
DOI：10.1007/s11633-024-1531-3.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The reasoning chain generated by the large language models (LLMs) during the reasoning process is often susceptible to illusions that lead to incorrect reasoning steps. Such misleading intermediate reasoning steps may trigger a series of errors. This phenomenon can be alleviated by using validation methods to obtain feedback and adjust the reasoning process, similar to the human reflective process. In this paper, we propose a collaborative reasoning framework for mathematical reasoning called CRMR, where a generator is responsible for generating structured intermediate reasoning and a verifier provides detailed feedback on each step of the reasoning. In particular, we formulate a rigorous form of structured intermediate reasoning called step-by-step rationale (SSR). We evaluated the CRMR framework not only on mathematical word problems but also conducted experiments using open-source and closed-source models with different parameter sizes independently. The results show that our method fully exploits the inference capabilities of the models and achieves significant results on the dataset compared to a single model.
Keywords:
Large language models
prompt engineering
natural language processing
cooperative reasoning
multi-step reasoning
Journal
IF:
8.7
Papers:
304
Citations:
882
Organization
Cited Papers
.png){kind=icon}
ShareSave
ShareSaveShareSave
View more