Aditya-Ranjan1234/mindgames_NeurIPS2025: Revac_8 Initial Release – NeurIPS 2025 Winning Agent
Fuente:
Zenodo
Salvato in:
| Autori principali: | , |
|---|---|
| Natura: | Recurso digital |
| Pubblicazione: |
Zenodo
2026
|
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866901664663863296 |
|---|---|
| author | mihiraryaa Avinash Anish |
| author_facet | mihiraryaa Avinash Anish |
| contents | <h1> Revac: Social Deduction Reasoning Agent</h1> <p> <strong>1st Place – MindGames Arena, NeurIPS 2025 (Social Deduction Track, Open Division)</strong></p> <h2> Overview</h2> <p>Revac is a multi-agent AI system designed for social deduction games such as Mafia, where agents must reason under uncertainty, interpret incomplete and deceptive information, and make strategic decisions through natural language interaction.</p> <p>Unlike traditional game-playing AI, Revac focuses on inference under partial observability, modeling deception and trust, and adaptive communication strategies. The system evolved from a simple reasoning pipeline into a modular, multi-stage architecture integrating memory, social reasoning, and dynamic communication.</p> <h2> Key Contributions</h2> <h3> Persistent Memory Module</h3> <p>Maintains long-term player profiles, tracks behavior and interactions across turns, and enables context-aware reasoning.</p> <h3> Social Alignment Graph (SAG)</h3> <p>A structured graph representation of accusations, defenses, and role claims used to detect collusion, alliances, and group pressure dynamics.</p> <h3> Dynamic Tone Selector (DTS)</h3> <p>Adapts communication strategy dynamically (aggressive, defensive, logical, contrarian) to improve persuasion and survivability.</p> <h2> Architecture Evolution</h2> <p>Revac → Revac2 → Revac8</p> <h3>Final Pipeline (Revac8)</h3> <p>Observation → Memory Update → Reviewer → Tone Selector → Action Agent → Output</p> <h2> Agent Variants</h2> <ul> <li>RevacAgent: Baseline two-stage reasoning (Reviewer → Action)</li> <li>Revac2Agent: Adds persistent memory and social tracking</li> <li>Revac8Agent: Full system with tone adaptation and advanced reasoning</li> <li>HumanAgent: Manual interface for testing</li> </ul> <h2> Environment: Secret Mafia</h2> <p>A turn-based social deduction environment with partial observability and asymmetric roles.</p> <h3>Roles</h3> <ul> <li>Village (Villagers, Doctor, Detective)</li> <li>Mafia</li> </ul> <h3>Phases</h3> <ul> <li>Night: Hidden actions (kill, protect, investigate)</li> <li>Day: Discussion, reasoning, voting</li> </ul> <h3>Challenges</h3> <ul> <li>Partial observability</li> <li>Asymmetric information</li> <li>Intentional deception</li> </ul> <h2> Performance</h2> <p>Final Standings (TrueSkill Rating):</p> <ul> <li> Revac_8 — 13.9</li> <li> Fractal Agent — 7.8</li> <li> Fractal Variant — 4.7</li> </ul> <p>Revac_8 significantly outperforms competing agents, demonstrating the effectiveness of structured memory and adaptive communication.</p> <h2>⚙️ Tech Stack</h2> <ul> <li>Python 3.8+</li> <li>LLM APIs (OpenAI / Groq)</li> <li>Modular agent pipeline</li> <li>Graph-based reasoning</li> <li>Prompt-engineered multi-stage inference</li> </ul> <h2>▶️ Usage</h2> <h3>Install</h3> <p>pip install -r src/requirements.txt</p> <h3>Run Simulation</h3> <p>python src/offline_play.py</p> <h3>Benchmark</h3> <p>python -m benchmark.test</p> <h2> Project Structure</h2> <p>src/ ├── agents/ # Agent implementations ├── prompts/ # Prompt templates ├── config.py # Configuration ├── offline_play.py └── utils.py</p> <p>envs/ # Game environment benchmark/ # Evaluation suite</p> <h2> Research Significance</h2> <p>Revac demonstrates that LLM-based agents can effectively operate in deceptive, multi-agent environments by combining memory-driven reasoning, social interaction modeling, and adaptive communication.</p> <p>This has implications for multi-agent systems, AI alignment, human-AI interaction, and strategic reasoning under uncertainty.</p> <h2> License</h2> <p>MIT License</p> <h2> Acknowledgments</h2> <p>Developed for MindGames Arena – NeurIPS 2025.</p> <h2> Future Work</h2> <ul> <li>Improved deception detection</li> <li>Multi-game generalization</li> <li>Learning-based strategy optimization</li> <li>Real-time human-agent interaction</li> </ul> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19555330 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Aditya-Ranjan1234/mindgames_NeurIPS2025: Revac_8 Initial Release – NeurIPS 2025 Winning Agent mihiraryaa Avinash Anish <h1> Revac: Social Deduction Reasoning Agent</h1> <p> <strong>1st Place – MindGames Arena, NeurIPS 2025 (Social Deduction Track, Open Division)</strong></p> <h2> Overview</h2> <p>Revac is a multi-agent AI system designed for social deduction games such as Mafia, where agents must reason under uncertainty, interpret incomplete and deceptive information, and make strategic decisions through natural language interaction.</p> <p>Unlike traditional game-playing AI, Revac focuses on inference under partial observability, modeling deception and trust, and adaptive communication strategies. The system evolved from a simple reasoning pipeline into a modular, multi-stage architecture integrating memory, social reasoning, and dynamic communication.</p> <h2> Key Contributions</h2> <h3> Persistent Memory Module</h3> <p>Maintains long-term player profiles, tracks behavior and interactions across turns, and enables context-aware reasoning.</p> <h3> Social Alignment Graph (SAG)</h3> <p>A structured graph representation of accusations, defenses, and role claims used to detect collusion, alliances, and group pressure dynamics.</p> <h3> Dynamic Tone Selector (DTS)</h3> <p>Adapts communication strategy dynamically (aggressive, defensive, logical, contrarian) to improve persuasion and survivability.</p> <h2> Architecture Evolution</h2> <p>Revac → Revac2 → Revac8</p> <h3>Final Pipeline (Revac8)</h3> <p>Observation → Memory Update → Reviewer → Tone Selector → Action Agent → Output</p> <h2> Agent Variants</h2> <ul> <li>RevacAgent: Baseline two-stage reasoning (Reviewer → Action)</li> <li>Revac2Agent: Adds persistent memory and social tracking</li> <li>Revac8Agent: Full system with tone adaptation and advanced reasoning</li> <li>HumanAgent: Manual interface for testing</li> </ul> <h2> Environment: Secret Mafia</h2> <p>A turn-based social deduction environment with partial observability and asymmetric roles.</p> <h3>Roles</h3> <ul> <li>Village (Villagers, Doctor, Detective)</li> <li>Mafia</li> </ul> <h3>Phases</h3> <ul> <li>Night: Hidden actions (kill, protect, investigate)</li> <li>Day: Discussion, reasoning, voting</li> </ul> <h3>Challenges</h3> <ul> <li>Partial observability</li> <li>Asymmetric information</li> <li>Intentional deception</li> </ul> <h2> Performance</h2> <p>Final Standings (TrueSkill Rating):</p> <ul> <li> Revac_8 — 13.9</li> <li> Fractal Agent — 7.8</li> <li> Fractal Variant — 4.7</li> </ul> <p>Revac_8 significantly outperforms competing agents, demonstrating the effectiveness of structured memory and adaptive communication.</p> <h2>⚙️ Tech Stack</h2> <ul> <li>Python 3.8+</li> <li>LLM APIs (OpenAI / Groq)</li> <li>Modular agent pipeline</li> <li>Graph-based reasoning</li> <li>Prompt-engineered multi-stage inference</li> </ul> <h2>▶️ Usage</h2> <h3>Install</h3> <p>pip install -r src/requirements.txt</p> <h3>Run Simulation</h3> <p>python src/offline_play.py</p> <h3>Benchmark</h3> <p>python -m benchmark.test</p> <h2> Project Structure</h2> <p>src/ ├── agents/ # Agent implementations ├── prompts/ # Prompt templates ├── config.py # Configuration ├── offline_play.py └── utils.py</p> <p>envs/ # Game environment benchmark/ # Evaluation suite</p> <h2> Research Significance</h2> <p>Revac demonstrates that LLM-based agents can effectively operate in deceptive, multi-agent environments by combining memory-driven reasoning, social interaction modeling, and adaptive communication.</p> <p>This has implications for multi-agent systems, AI alignment, human-AI interaction, and strategic reasoning under uncertainty.</p> <h2> License</h2> <p>MIT License</p> <h2> Acknowledgments</h2> <p>Developed for MindGames Arena – NeurIPS 2025.</p> <h2> Future Work</h2> <ul> <li>Improved deception detection</li> <li>Multi-game generalization</li> <li>Learning-based strategy optimization</li> <li>Real-time human-agent interaction</li> </ul> |
| title | Aditya-Ranjan1234/mindgames_NeurIPS2025: Revac_8 Initial Release – NeurIPS 2025 Winning Agent |
| url | https://doi.org/10.5281/zenodo.19555330 |