Beyond Nash Equilibrium: Bounded Rationality of LLMs and humans in Strategic Decision-making

Kehan Zheng, Jinfeng Zhou, Jian Guan, Wei Wu, Minlie Huang, Hongning Wang

Keywords: Large Language Models, Strategic Decision-Making, Bounded Rationality, Game Theory

Abstract: Large language models are increasingly used in strategic decision-making settings, yet evidence shows that, like humans, they often deviate from full rationality. In this study, we compare LLMs and humans using experimental paradigms directly adopted from behavioral game-theory research. To ensure a precise comparison, we experiment with a finite action space, a closed-form Nash equilibrium benchmark, and well-documented human deviations from that equilibrium. These criteria naturally yield two complementary game families: a zero-sum family exemplified by Rock-Paper-Scissors and a non-zero-sum family exemplified by the Prisoner’s Dilemma, whose parameters we systematically vary to form a broader task space. By placing LLMs in identical experimental conditions, we evaluate whether their behaviors exhibit the bounded rationality characteristic of humans. Our findings show that LLMs reproduce similar human heuristics, such as outcome‑based strategy switching and increased cooperation when future interaction is possible, but LLMs follow these patterns more rigidly and demonstrate weaker sensitivity to the dynamic changes in the game environment. Model-level analyses reveal distinctive architectural signatures in strategic behavior, yet reasoning-enhanced models still falter in dynamic, opponent-adaptive settings, showing that a long chain of thought alone does not guarantee full rationality. These results indicate that current LLMs capture only a partial form of human-like bounded rationality and highlight the need for training methods that encourage flexible opponent modeling and stronger context awareness.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 15381