Welcome to Game Reasoning Arena’s documentation!

Game Reasoning Arena is a research platform for training and evaluating AI agents in games using Large Language Models and reinforcement learning techniques.

Project Overview

Game Reasoning Arena provides a comprehensive framework for:

  • Multi-Agent Testing: Compare LLMs vs Random, LLM vs LLM, and Self-play scenarios

  • Multiple Game Types: Strategy games, poker variants, cooperation games, and zero-sum games

  • Flexible Backends: Support for API-based (LiteLLM) and local (vLLM) inference

  • Cross-Provider Compatibility: Mix different LLM providers within the same game

  • Extensible Architecture: Easy to add new games, agents, and analysis tools

Quick Start

The framework is designed around a few key concepts:

  • Environments: Game simulations built on OpenSpiel

  • Agents: AI players including LLMs, random agents, and human players

  • Backends: Flexible LLM inference systems (local and API-based)

  • Analysis Tools: Post-game reasoning analysis and visualization

See the Installation guide to get started, or explore the API Reference for detailed project structure information.

Available Games

  • tic_tac_toe - Classic 3×3 grid strategy game

  • connect_four - Drop pieces to connect four in a row

  • kuhn_poker - Simple poker variant with hidden information

  • prisoners_dilemma - Cooperation vs defection scenarios

  • matching_pennies - Zero-sum matching game

  • matrix_rps - Rock-paper-scissors in matrix form

  • hex - Long strategy game on a hexagonal grid