Reproduce “Suspicion-Agent: Playing Imperfect Information Games with Theory of Mind Aware GPT-4” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiaxian Guo, Bo Yang, Paul Yoo, Bill Yuchen Lin, Yusuke Iwasawa, Yutaka Matsuo
This paper presents Suspicion-Agent, a GPT-4-based framework that plays imperfect information games zero-shot, without specific training. It integrates progressive levels of Theory of Mind, enabling complex strategies like bluffing and dynamic adaptation, outperforming many traditional learning-based AI methods.
Key insights
- Large Language Models, specifically GPT-4, exhibit emergent higher-order Theory of Mind capabilities that enable sophisticated strategic play like bluffing and dynamic adaptation in imperfect information games.
- A modular, prompt-engineered framework can enable GPT-4 to play diverse imperfect information games in a zero-shot manner, without any game-specific training or examples.
- The advanced reasoning and instruction following of GPT-4 are crucial, as earlier versions like GPT-3.5 perform significantly worse in these complex strategic environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.