"Game-theoretic LLM: Agent Workflow for Negotiation Games"

LLMs can now play chess with game theory, making smarter moves in strategic decisions

And Game theory workflows transform LLMs from random players to strategic thinkers

This paper investigates LLMs' rationality in strategic decision-making, particularly in game theory contexts. The research reveals LLMs often deviate from optimal strategies in complex scenarios but demonstrates how specialized game-theoretic workflows can significantly improve their performance.

-----

https://arxiv.org/abs/2411.05990

🎯 Original Problem:

LLMs struggle with rational decision-making in strategic scenarios, especially when dealing with complex payoff matrices or sequential decision trees. Their performance deteriorates in games requiring coordination and strategic thinking.

-----

🔧 Solution in this Paper:

→ The researchers designed multiple game-theoretic workflows to guide LLM reasoning and decision-making processes

→ These workflows incorporate principles like Dominant Strategy Search and Backward Induction to help compute Nash Equilibria

→ The solution integrates classic game theory strategies into LLM agent workflows for both complete and incomplete information games

→ A structured approach helps LLMs make rational choices under uncertainty through Bayesian belief updating

-----

💡 Key Insights:

→ LLMs show significant improvement in identifying optimal strategies when using structured workflows

→ The decision to use workflows itself becomes a game-theoretic consideration

→ There are clear synergies between improving strategic decision-making and reducing exploitation risks

-----

📊 Results:

→ Workflow adoption significantly enhanced LLM rationality in game-theoretic tasks

→ LLMs achieved near-optimal allocations in negotiation scenarios

→ Strategic decision-making improved across both complete and incomplete information games