🏆 1st Place(Team Phoebus) - MindGames Arena NeurIPS 2025 (Social Deduction Track, Open Division)
Social deduction games such as Mafia present a unique AI challenge: players must reason under uncertainty, interpret incomplete and intentionally misleading information, evaluate human-like communication, and make strategic elimination decisions. Unlike deterministic board games, success depends not on perfect information or brute-force search, but on inference, memory, and adaptability in the presence of deception.
Revac is an AI agent designed for multi-agent social deduction games, built on a modular, multi-stage reasoning architecture. The agent evolved from a simple two-stage reasoning system (Revac) to a sophisticated multi-module architecture (Revac_8) that integrates memory-based player profiling, social-graph analysis of accusations and defenses, and dynamic tone selection for adaptive communication.
This repository contains the implementation of the Revac agent family, developed for the MindGames Arena competition, where Revac_8 achieved first place in the Social Deduction track. The agent's success highlights the critical role of structured memory and adaptive communication in achieving competitive performance in high-stakes social environments.
📖 For detailed technical documentation, architecture deep-dives, and memory system explanations, see
src/agents/README.md
Revac_8 secured first place in the Open Division of the Social Deduction Track at MindGames Arena NeurIPS 2025.
| Rank | Agent Name | TrueSkill Rating |
|---|---|---|
| 🥇 1st | Revac_8 | 13.9 |
| 🥈 2nd | Fractal_SecretMafia_Agent_round2_v25 | 7.8 |
| 🥉 3rd | Fractal_SecretMafia_Agent_round2_v14u | 4.7 |
The competition evaluated agents across multiple games of Secret Mafia, a turn-based social deduction environment with partial observability and asymmetric roles (Village-aligned vs Mafia).
- Social Alignment Graph (SAG): Structured graph representation of social interactions (accusations, defenses, role claims) enabling collusion detection and group pressure analysis
- Persistent Memory Module: Long-term tracking of player profiles and social relationships across all game turns
- Dynamic Tone Selector (DTS): Adaptive communication strategies that adjust tone based on strategic context (aggressive, withdrawing, logically anchoring, contrarian)
The Revac agent family evolved through three iterations, each adding new capabilities:
Revac → Revac2 (+ Memory) → Revac8 (+ Tone Selection)
All agents use a modular pipeline architecture. The final Revac8Agent pipeline:
Observation → Memory Update → Reviewer → Tone Selector → Action Agent → Output
💡 See
src/agents/README.mdfor detailed architecture documentation, component explanations, memory system deep-dive.
- Python 3.8+
- OpenAI API key or Groq API key
- Clone the repository
git clone https://github.com/mihiraryaa/mindgames_NeurIPS2025.git
cd mindgames_NeurIPS2025- Create and activate virtual environment
python -m venv venv
# Windows
venv\Scripts\activate
# Linux/Mac
source venv/bin/activate- Install dependencies
pip install -r src/requirements.txt- Configure API keys
Create a .env file in the project root:
OPENAI_API_KEY=your_openai_key_here
GROQ_API_KEY=your_groq_key_here
# Optional: Supabase logging (set ENABLE_LOGGING=True in config.py)
SUPABASE_URL=your_supabase_url
SUPABASE_KEY=your_supabase_keypython src/offline_play.pyThis will run a local Secret Mafia game with multiple Revac agents.
from src.agents import Revac8Agent
# Initialize agent
agent = Revac8Agent("gpt-5-mini", "openai")
# Use in game loop
observation = "Game observation string"
action = agent(observation)python -m benchmark.test📚 For more examples including TextArena integration, different model providers, and custom agent creation, see
src/agents/README.md
mindgames_NeurIPS2025/
├── src/
│ ├── agents/ # Agent implementations (Revac, Revac2, Revac8)
│ │ ├── base.py # Abstract Agent class
│ │ ├── revac.py # Two-stage reasoning agent
│ │ ├── revac2_1.py # Memory-enhanced agent
│ │ ├── revac8.py # Tone-adaptive agent
│ │ ├── judge_agent.py # Benchmark evaluation agent
│ │ ├── schemas.py # Pydantic models (Memory, SocialRelation, etc.)
│ │ └── README.md # Detailed agent documentation
│ │
│ ├── prompts/ # System prompts for each agent variant
│ │ ├── revac_prompts.py
│ │ ├── revac2_1_prompts.py
│ │ └── revac8_prompts.py
│ │
│ ├── config.py # Environment configuration and API keys
│ ├── utils.py # Helper functions (get_model, logging)
│ ├── offline_play.py # Main script for local game simulation
│ └── requirements.txt # Python dependencies
│
├── envs/ # Custom TextArena environment configurations
│ ├── SecretMafia/ # Social deduction game environment
│ └── README.md # Environment documentation
│
├── benchmark/ # Benchmark test suite
│ ├── test.py # Benchmark runner
│ ├── cases/ # Test cases (challenging scenarios)
│
│
├── LICENSE # MIT License
└── README.md # This file
This project includes four agent implementations:
- RevacAgent - Baseline two-stage reasoning architecture (Reviewer → Action)
- Revac2Agent - Adds persistent memory with Player Profiles and Social Alignment Graph
- Revac8Agent - Extends Revac2 with dynamic tone selection for adaptive communication
- HumanAgent - Manual control interface for testing and debugging
📋 For detailed comparison table, architecture diagrams, memory system explanations, and usage examples, see
src/agents/README.md
The agents are evaluated in a turn-based social deduction game based on Mafia:
- Roles: Village-aligned (Villagers, Doctor, Detective) vs Mafia
- Phases:
- Night: Private actions (Mafia kills, Doctor protects, Detective investigates)
- Day: Open discussion, voting, and elimination
- Objective:
- Village: Eliminate all Mafia members
- Mafia: Achieve numerical parity with Village
- Challenge: Partial observability, asymmetric information, intentional deception
This project is licensed under the MIT License - see the LICENSE file for details.
Built for the MindGames Arena NeurIPS 2025 challenge. Special thanks to the oraganizing team for providing the game environments and competition infrastructure.