- Evaluating the Cooperative Potential of LLMs
- Project Year:
2024
- REU Student (s):
Tymur Kotkov | Charles University (Prague, Czech Republic)
- Student 1 Institution:
Charles University (Prague, Czech Republic)
- Project Mentor:
Xintong Wang
- Project Mentor Area:
Computer Science
- Project Abstract:
Large Language Models (LLMs) are increasingly pivotal in enhancing human-AI interactions, yet their behavior in competitive environments remains underexplored. This study investigates whether LLMs can develop cooperative strategies in the Prisoner's Dilemma, knowing only the game rules and receiving no external decision-making support. We simulate an Axelrod's Tournament with a limited set of predefined strategies and one LLM agent, assessing its behavior based on successful strategy characteristics and tournament outcomes. Our findings reveal that the LLM adapts its decision-making process to each opponent, demonstrating an ability to balance cooperation and competition effectively. This adaptability suggests significant potential for deploying LLMs in complex, dynamic environments such as economics, diplomacy, and social governance. While further validation in real-world scenarios is necessary, these results provide a promising foundation for understanding how LLMs make choices within different environments and contexts.
