Background

The prisoner’s dilemma refers to a scenario in which two conspirators in a crime are arrested and imprisoned. There is only enough evidence to convict them on a minor charge with a sentence of 5 years. The police offer each prisoner a deal: if she testifies against the other prisoner, the other prisoner will receive the full 10-year sentence for the more serious crime while she goes free. If both prisoners agree to testify, they will each serve a reduced sentence of 8 years. The prisoners cannot communicate with each other. They ensure the best joint outcome if neither of them cooperates with the police, but a prisoner risks the worst individual outcome if she stays quiet and the other prisoner testifies.

The prisoner’s dilemma is a key problem in game theory, and researchers often study the iterated version of the game, in which the same two individuals play multiple rounds of the prisoner’s dilemma. Each player can try to guess whether the other is likely to cooperate or not based on their prior behavior. The iterated prisoner’s dilemma served as a lens for analyzing the nuclear arms race during the Cold War, and it has also provided insights in biology and economics. Computer simulations allow experimenters to see how different strategies perform in the iterated prisoner’s dilemma. Recent work has expanded these simulations to large language models, and researchers have discovered different LLMs choose different strategies, with some being more aggressive than others.

Research Goals

This project seeks to replicate prior findings regarding LLMs’ performance in iterated prisoner’s dilemma tournaments, beginning with a systematic literature review of relevant prior work on the topic. We will also include LLMs not previously studied and compare the performance of LLMs from different countries. LLMs’ post-training may vary depending on the cultural and legal context in which they were developed, and this may affect the strategies they choose when playing the iterated prisoner’s dilemma.

Skills Needed

Applicants should be willing to read research papers, including papers from disciplines other than computer science. Applicants should have enough programming experience to interact with LLMs via an API; Python will likely be the most useful programming language for this task. Knowledge of game theory would be helpful but is not required.

Skills Gained

Experiment design; systematic literature review; academic paper writing