The article presents LINC, a neurosymbolic approach for logical reasoning in artificial intelligence. LINC combines Large Language Models (LLMs) with first-order logic provers. In this method, the LLM acts as a semantic parser, translating premises and conclusions from natural language to expressions in first-order logic. These expressions are then processed by an external theorem prover, which symbolically performs deductive inference. The study finds that LINC significantly improves performance on logical reasoning tasks, even outperforming models like GPT-3.5 and GPT-4 with Chain-of-Thought prompting. The authors suggest that LINC’s approach to logical reasoning over natural language could be beneficial when used alongside symbolic provers.
Publication date: 24 Oct 2023
Project Page: https://github.com/benlipkin/linc
Paper: https://arxiv.org/pdf/2310.15164
