jedell / spyfall

Spyfall Environment

Overview

This project develops a multi-agent social deduction game for AI research, inspired by Facebook Research's Cicero. The goal of this environment is to provide a framework for evaluating the ability of agents to craft and detect deceptive behavior in a social setting. For related work I found interesting, see:

• Human-level play in the game of Diplomacy by combining language models with strategic reasoning
• How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions
• Hoodwinked: Deception and Cooperation in a Text-Based Game for Language Models

Features

• Multi-agent gameplay
• Dialogue interactions to deduce roles and strategies.
• Integration with OpenAI's GPT models for generating responses.
• Interactive simulation of environment. python -m spyfall.environment.interactive

Current Implementation

• Dialogue agents that respond based on game state and past interactions.
• Game environment management for roles, locations, and actions.

Future Directions

• Enhance dialogue and action modules for complex reasoning based on dialogue.
• Lie detection and suspicion modules.
• Develop a MARL model that can play the game.

Contributions

Feel free to clone the repository, make improvements, and submit a pull request.