Reasoning Capabilities and Invariability of Large Language Models

TL;DR

This work tackles the robustness of large language models’ reasoning to prompt variations by introducing a 432-question, geometry-based benchmark designed for shallow logical reasoning. It systematically evaluates 24 LLMs under zero-shot, few-shot, and chain-of-thought prompting (across 22 models for CoT) and analyzes invariability using four question variants. The findings show GPT-4 achieves the top zero-shot performance at $72.30\%$, yet most models cluster near a $51\%$ baseline, while chain-of-thought prompts yield unpredictable gains depending on the timing of rationale; invariability analyses reveal stable behavior across minor linguistic changes and across models. Collectively, the results indicate that current LLMs lack robust, language-invariant reasoning under simple logical tasks and point to future directions beyond prompting, such as structured knowledge integration and misinformation-resistant reasoning.

Abstract

Large Language Models (LLMs) have shown remarkable capabilities in manipulating natural language across multiple applications, but their ability to handle simple reasoning tasks is often questioned. In this work, we aim to provide a comprehensive analysis of LLMs' reasoning competence, specifically focusing on their prompt dependency. In particular, we introduce a new benchmark dataset with a series of simple reasoning questions demanding shallow logical reasoning. Aligned with cognitive psychology standards, the questions are confined to a basic domain revolving around geometric figures, ensuring that responses are independent of any pre-existing intuition about the world and rely solely on deduction. An empirical analysis involving zero-shot and few-shot prompting across 24 LLMs of different sizes reveals that, while LLMs with over 70 billion parameters perform better in the zero-shot setting, there is still a large room for improvement. An additional test with chain-of-thought prompting over 22 LLMs shows that this additional prompt can aid or damage the performance of models, depending on whether the rationale is required before or after the answer.