Rationality of Thought Improves Reasoning in Large Language Models

While the capabilities of large language models (LLMs) have been progressively advanced, their competence in addressing intricate reasoning tasks remains inadequate, primarily due to their insufficient cognitive capabilities. To explore the cognitive proficiency of models like GPT-4, we turn to methodologies from cognitive psychology: cognitive abilities reflect rational thinking skills, and cognitive bias tasks are often used to assess rational thinking levels. In this paper, we develop a cognitive bias dataset to measure the rational thinking and cognitive levels of LLMs. Our observations indicate that GPT-4, akin to humans, exhibits limitations in its rational thinking ability. We propose a new method, “Rationality of Thought” (RoT), to prompt LLMs into a rational thinking process during task execution. This method significantly improves the accuracy of GPT-4 on the cognitive bias task by 18.7%. Cognitive capacity is also essential for tackling complex issues, therefore, we implement RoT across various reasoning tasks. Using only a zero-shot setting, RoT outperforms inference enhancement techniques such as CoT using zero-shot, such as SVAMP(+1.8),AQUA-RAT (+6.0), ARC-c (+4.1) ,ARC-e(+3.9) in multiple arithmetic and common sense reasoning tasks. Our empirical evaluation shows that RoT helps LLMs elevate their cognitive capabilities through rational thinking, thereby becoming more adept at navigating complex reasoning tasks.