Core Concepts
The core message of this article is to propose a novel reasoning framework, named Generalizable and Faithful Reasoner (GFaiR), which introduces the paradigm of resolution refutation to improve the completeness and faithfulness of logic reasoning over natural language.
Abstract
The article presents a novel reasoning framework called GFaiR that aims to address the limitations of previous transformer-based reasoning systems. The key insights are:
Previous LLMs-based reasoning systems suffer from theoretical incompleteness, which restricts their ability to handle complex reasoning problems. To address this, GFaiR introduces the paradigm of resolution refutation, which has the capability to solve all first-order logic reasoning problems.
GFaiR consists of five key modules: a converter, a pre-selector, a post-selector, a knowledge composer, and a verifier. The converter transforms the natural language theory and hypothesis into a format suitable for resolution. The pre-selector and post-selector select relevant theories for the knowledge composer to apply resolution. The verifier ensures the selected theories can form a valid theory pair for resolution, improving the faithfulness of the reasoning process.
Experimental results show that GFaiR outperforms previous methods on complex reasoning scenarios while maintaining performance on simple scenarios. GFaiR also demonstrates strong zero-shot generalization abilities and faithfulness to its reasoning process.
The article also evaluates GFaiR on the natural language satisfiability (NLSAT) task, which requires reasoning solely based on rules without any facts. GFaiR outperforms the baseline methods on this more challenging task, further demonstrating its capability in handling complex reasoning scenarios.