Sunday 06 April 2025
Researchers have been exploring the capabilities of large language models (LLMs) in various scenarios, including game theory. A recent study delves into the performance of LLMs in playing two classic games: Rock-Paper-Scissors and Prisoner’s Dilemma.
The first game, Rock-Paper-Scissors, is a well-known example of a zero-sum game, where one player’s gain is equal to the other player’s loss. In this study, the researchers found that LLMs were unable to reliably achieve uniform distribution in their move selection, even after repeated games. This suggests that the models are not truly random, but rather tend towards a pattern.
In contrast, when playing Prisoner’s Dilemma, a game where two players must make cooperative or competitive decisions, the LLMs showed a surprising ability to adapt and learn from experience. However, this adaptation came at a cost: the models often converged on the Nash equilibrium, a strategy that leads to suboptimal outcomes for both players.
The researchers also experimented with different prompts, designed to influence the models’ behavior. Some prompts encouraged cooperation, while others emphasized competition. The results showed that the LLMs were sensitive to these cues and adjusted their strategies accordingly.
One of the most intriguing findings was the emergence of loss-averse behavior in the repeated Prisoner’s Dilemma games. The models seemed to prioritize avoiding losses over maximizing gains, leading to a stalemate rather than a competitive outcome. This raises questions about the potential applications of LLMs in real-world scenarios, where cooperation and competition are often intertwined.
The study highlights the complexities of large language models and their limitations in certain situations. While they can demonstrate impressive abilities in some domains, they are not omniscient or infallible. The results also underscore the importance of understanding the underlying mechanisms driving LLM behavior, as well as designing effective prompts to elicit desired responses.
As researchers continue to explore the capabilities and limitations of large language models, we may uncover new insights into their potential applications in fields such as finance, education, and healthcare. For now, this study serves as a reminder that even the most advanced AI systems are not immune to the complexities of human decision-making.
Cite this article: “LLMs Game Theory: Insights into the Strategic Limitations of Large Language Models”, The Science Archive, 2025.
Large Language Models, Game Theory, Rock-Paper-Scissors, Prisoner’S Dilemma, Zero-Sum Games, Nash Equilibrium, Loss Aversion, Cooperation, Competition, Ai Systems
