LLMs Outperform Humans in Strategic Game Behavior Study
A new study using AlphaEvolve, a program discovery tool, analyzed strategic behavior in iterated rock-paper-scissors and found that frontier Large Language Models (LLMs) exhibit deeper strategic reasoning than humans. The research, published on arXiv, employs behavioral game theory to model and compare decision-making, revealing structural differences between human and LLM behavior in social and strategic contexts.
Key facts
- LLMs are increasingly deployed in social and strategic scenarios
- Behavioral game theory provides a framework for analyzing behavior
- Existing models do not fully capture idiosyncratic behavior of humans or LLMs
- AlphaEvolve is a cutting-edge program discovery tool
- AlphaEvolve directly discovers interpretable models from data
- Analysis was conducted on iterated rock-paper-scissors
- Frontier LLMs can be capable of deeper strategic behavior than humans
- Results provide foundation for understanding structural differences in strategic interactions
Entities
Institutions
- arXiv