Reproduce “Richelieu: Self-Evolving LLM-Based Agents for AI Diplomacy” 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

Zhenyu Guan, Xiangyu Kong, Fangwei Zhong, Yizhou Wang

Richelieu introduces a self-evolving AI diplomacy agent that learns complex social and strategic behaviors through self-play alone, eliminating the need for human demonstration data. It achieves a 6.83% win rate in Press Diplomacy against the human-data-dependent Cicero's 6.03%, demonstrating improved performance and generalizability across various large language models.

Key insights

  • An LLM-based agent can master complex multi-agent social games like Diplomacy solely through self-play, without requiring any human demonstration data, overcoming a major limitation of previous approaches.
  • Integrating modules for social reasoning (modeling relationships and inferring intentions), strategic planning with reflection, and goal-oriented negotiation enables LLMs to achieve human-like diplomatic capabilities.
  • The framework is robust and generalizable, enhancing the performance of various underlying LLMs (e.g., GPT-4.0, Llama 3) and showing potential for application in other complex social deduction games.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.