GPT-3.5 vs GPT-4: Evaluating ChatGPT’s Reasoning Performance in Zero-shot Learning

Discover the first version of our scientific publication "GPT-3.5 vs GPT-4: Evaluating ChatGPT's Reasoning Performance in Zero-shot Learning" published in arxiv, a widely recognized platform for sharing preprints and scientific articles. This article is currently undergoing a rigorous review process.

Thanks to the Novelis research team - including Jessica López Espejel, Mahaman Sanoussi Yahaya Alassan, El Mehdi Chouham, El Hassane Ettifouri, Walid Dahhane - for their know-how and expertise.

Abstract

“Large Language Models (LLMs) have exhibited remarkable performance on various Natural Language Processing (NLP) tasks. However, there is a current hot debate regarding their reasoning capacity. In this paper, we examine the performance of GPT-3.5 and GPT-4 models, by performing a thorough technical evaluation on different reasoning tasks across eleven distinct datasets. Our findings show that GPT-4 outperforms GPT-3.5 in zero-shot learning throughout almost all evaluated tasks. In addition, we note that both models exhibit limited performance in Inductive, Mathematical, and Multi-hop Reasoning Tasks. While it may seem intuitive that the GPT-4 model would outperform GPT-3.5 given its size and efficiency in various NLP tasks, our paper offers empirical evidence to support this claim. We provide a detailed and comprehensive analysis of the results from both models to further support our findings. In addition, we propose a set of engineered prompts that improves performance of both models on zero-shot learning.”

arXiv is an open archive of electronic preprints of scientific articles in various technical fields, such as physics, mathematics, computer science and more, freely accessible via the Internet.

Cornell University is a private American university located primarily in the city of Ithaca, New York.