Large language models instantiate evolutionarily robust strategies of cooperation

Area of Science:

• Artificial Intelligence
• Game Theory
• Behavioral Economics

Background:

• Large language models (LLMs) are increasingly used to aid human decision-making, raising concerns about their social behavior.
• The repeated prisoner's dilemma is a key model for studying reciprocal cooperation and social strategies.

Purpose of the Study:

• To evaluate the social decision-making strategies of five leading LLMs in the repeated prisoner's dilemma.
• To assess LLM adaptability to game parameters and framing effects compared to theoretical predictions and human behavior.

Main Methods:

• LLMs were tested for cooperation propensity in neutral settings and adherence to game theory concepts (e.g., Nash equilibria).
• Strategies were analyzed under varying parameters (stopping probability, payoffs, known/unknown rounds) and different framings.
• Tournaments and direct LLM interactions were conducted to observe emergent strategies.

Main Results:

• LLMs demonstrated competence in several tasks but lacked full consistency across all tested conditions.
• Adaptations to parameter changes and framing effects were not always aligned with evolutionary game theory or human experimental results.
• LLM strategies showed varied performance in direct interactions and tournaments.

Conclusions:

• Current LLMs exhibit a developing capacity for reciprocal cooperation, but their social decision-making is not fully robust or predictable.
• Further research is needed to understand and refine LLM social behavior for reliable integration into decision-support systems.