Language Models can be Logical Solvers

逻辑推理是人类智能的基本方面，也是问题解决和决策制定等任务的关键组成部分。最近的进展使得大型语言模型（LLM）有可能展现出推理能力，但复杂的逻辑推理仍然是一个挑战。目前最先进的求解器增强语言模型使用LLM先解析自然语言逻辑问题为符号表示，然后采用外部逻辑求解器接收符号表示并输出答案。尽管它们的表现令人印象深刻，但任何解析错误都将不可避免地导致外部逻辑求解器的执行失败，无法回答逻辑问题。本文介绍了一种新型语言模型LoGiPT，它直接模拟逻辑求解器的推理过程，并通过学习严格遵守求解器的语法和语法规则来绕过解析错误。LoGiPT在新构建的指令调整数据集上进行微调，该数据集源于揭示和完善演绎求解器的不可见推理过程。在两个公共演绎推理数据集上的实验结果表明，LoGiPT在竞争性LLM（如ChatGPT或GPT-4）上表现优于最先进的求解器增强语言模型和少量提示方法。

The paper aims to introduce LoGiPT, a language model that emulates the reasoning processes of logical solvers and bypasses parsing errors by learning strict adherence to solver syntax and grammar. The goal is to improve upon the state-of-the-art solver-augmented language models and few-shot prompting methods on competitive LLMs like ChatGPT or GPT-4 for deductive reasoning tasks.

The key idea of the paper is to directly emulate the reasoning processes of logical solvers in a language model, rather than relying on external solvers to interpret symbolic representations of natural language logical questions. LoGiPT is fine-tuned on a newly constructed instruction-tuning dataset derived from revealing and refining the invisible reasoning process of deductive solvers.

The paper demonstrates that LoGiPT outperforms state-of-the-art solver-augmented LMs and few-shot prompting methods on competitive LLMs like ChatGPT or GPT-4 for deductive reasoning tasks. The experimental results are based on two public deductive reasoning datasets. The paper also highlights the importance of strict adherence to solver syntax and grammar in order to bypass parsing errors. The authors provide an open-source implementation of LoGiPT and the instruction-tuning dataset for further research.

Recent related work in this field includes the use of solver-augmented language models for deductive reasoning tasks, as well as few-shot prompting methods for improving language models in general. Some relevant papers include 'Logical Natural Language Generation from Open-Domain Tables' by Yixin Nie et al. and 'Few-Shot Learning for Natural Language Processing: A Survey' by Mingda Chen et al.