LogicBench: Towards Systematic Evaluation of Logical Reasoning Ability of Large Language Models

TL;DR

LogicBench provides a comprehensive, isolated evaluation of large language models' logical reasoning across 25 patterns spanning propositional, first-order, and non-monotonic logics. The authors introduce a three-stage data-generation pipeline to produce natural-language contexts and two QA tasks (Binary QA and MCQA) anchored to single inference rules, enabling clear, metric-based comparisons and analysis of reasoning chains. Experiments across GPT-4, ChatGPT, Gemini, Llama-2, and Mistral reveal that current LLMs struggle with complex contexts, negations, and longer rules, though larger models perform better. The work also demonstrates that augmenting LogicBench (LogicBenchAug) can transfer gains to other logic datasets via fine-tuning (LogicT5) and improves performance on LogicNLI and FOLIO, suggesting practical avenues for enhancing logical reasoning in LLMs. Overall, LogicBench offers a rigorous benchmark and methodology to drive future improvements in LLM logical reasoning and invites extensions to additional rules and multilingual settings.

Abstract

Recently developed large language models (LLMs) have been shown to perform remarkably well on a wide range of language understanding tasks. But, can they really "reason" over the natural language? This question has been receiving significant research attention and many reasoning skills such as commonsense, numerical, and qualitative have been studied. However, the crucial skill pertaining to 'logical reasoning' has remained underexplored. Existing work investigating this reasoning ability of LLMs has focused only on a couple of inference rules (such as modus ponens and modus tollens) of propositional and first-order logic. Addressing the above limitation, we comprehensively evaluate the logical reasoning ability of LLMs on 25 different reasoning patterns spanning over propositional, first-order, and non-monotonic logics. To enable systematic evaluation, we introduce LogicBench, a natural language question-answering dataset focusing on the use of a single inference rule. We conduct detailed analysis with a range of LLMs such as GPT-4, ChatGPT, Gemini, Llama-2, and Mistral using chain-of-thought prompting. Experimental results show that existing LLMs do not fare well on LogicBench; especially, they struggle with instances involving complex reasoning and negations. Furthermore, they sometimes overlook contextual information necessary for reasoning to arrive at the correct conclusion. We believe that our work and findings facilitate future research for evaluating and enhancing the logical reasoning ability of LLMs. Data and code are available at https://github.com/Mihir3009/LogicBench.

Figures (5)

• Figure 1: Comprehensive representation of different inference rules and reasoning patterns covered by propositional, first-order, and non-monotonic logics. Exp. indicates Expectation
• Figure 2: Schematic representation of three-stage procedure for data creation. NL: Natural Language
• Figure 3: Schematic representation of prompt.
• Figure 4: Example prompt for Modus Tollens inference rule from PL.
• Figure 5: Word cloud of context present in the LB(eval)