Comparing Rationality Between Large Language Models and Humans: Insights and Open Questions
Dana Alsagheer, Rabimba Karanjai, Nour Diallo, Weidong Shi, Yang Lu, Suha Beydoun, Qiaoning Zhang
TL;DR
This study investigates how rationality in large language models (LLMs) compares with human performance, focusing on Reinforcement Learning from Human Feedback (RLHF) as a driver of improved decision-making in models like ChatGPT. It employs a battery of rationality tests (Wason selection task, conjunction fallacy, base rate neglect) across in-person and online human data, plus API-based assessments of ChatGPT (and Gemini), to quantify irrationality and alignment gaps. Key contributions include a comparative analysis of rationality across humans and LLMs, an examination of how human feedback influences LLM behavior, and a discussion of transparency, auditing, and open questions for responsible RLHF deployment. The work highlights the need for rigorous evaluation frameworks and governance mechanisms to safely harness LLMs in real-world decision contexts while acknowledging the role of human irrationality in feedback and model behavior.
Abstract
This paper delves into the dynamic landscape of artificial intelligence, specifically focusing on the burgeoning prominence of large language models (LLMs). We underscore the pivotal role of Reinforcement Learning from Human Feedback (RLHF) in augmenting LLMs' rationality and decision-making prowess. By meticulously examining the intricate relationship between human interaction and LLM behavior, we explore questions surrounding rationality and performance disparities between humans and LLMs, with particular attention to the Chat Generative Pre-trained Transformer. Our research employs comprehensive comparative analysis and delves into the inherent challenges of irrationality in LLMs, offering valuable insights and actionable strategies for enhancing their rationality. These findings hold significant implications for the widespread adoption of LLMs across diverse domains and applications, underscoring their potential to catalyze advancements in artificial intelligence.