Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Online Access: | https://doi.org/10.5281/zenodo.19556028 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <h1><strong>MindGames Arena NeurIPS 2025 — 1st Place, Social Deduction Track (Open Division)</strong></h1> <h2><strong>Summary</strong></h2> <p>Revac_8 is the final agent variant submitted to the MindGames Arena NeurIPS 2025 competition, where it achieved first place in the Social Deduction track with a TrueSkill rating of 13.9. This release includes the full agent family (Revac, Revac2, Revac8), game environment configurations, benchmark tooling, and accompanying paper.</p> <h2><strong>What's Included</strong></h2> <ul> <li><code>src/agents/</code> — All agent implementations: RevacAgent, Revac2Agent, Revac8Agent, HumanAgent</li> <li><code>src/prompts/</code> — System prompt sets for each agent variant</li> <li><code>src/offline_play.py</code> — Local game simulation runner</li> <li><code>envs/SecretMafia/</code> — Custom TextArena environment for Secret Mafia</li> <li><code>benchmark/</code> — Benchmark runner and test case suite</li> <li><code>paper.pdf</code> — Technical paper describing the architecture and results</li> </ul> <h2><strong>Agent Variants</strong></h2> <h3><strong>RevacAgent</strong></h3> <p>Baseline two-stage reasoning pipeline: Reviewer followed by Action Agent. No persistent state between turns.</p> <h3><strong>Revac2Agent</strong></h3> <p>Adds a persistent memory module with per-player profiling and a Social Alignment Graph (SAG) that tracks accusations, defenses, and role claims across the full game.</p> <h3><strong>Revac8Agent</strong></h3> <p>Extends Revac2 with a Dynamic Tone Selector (DTS) that adapts communication strategy based on strategic context. The full pipeline is: Observation -> Memory Update -> Reviewer -> Tone Selector -> Action Agent -> Output.</p> <h2><strong>Key Technical Contributions</strong></h2> <p><strong>Social Alignment Graph (SAG)</strong> A structured graph representation of social interactions across game turns. Edges encode accusation and defense events, enabling collusion detection and group pressure modeling.</p> <p><strong>Persistent Memory Module</strong> Long-term player profiles maintained across all turns, tracking behavioral patterns, stated role claims, and vote history.</p> <p><strong>Dynamic Tone Selector (DTS)</strong> Selects from a set of strategic communication tones — aggressive, withdrawing, logically anchoring, contrarian — based on the current game state and social graph analysis.</p> <h2><strong>Competition Results</strong></h2> <table> <tbody> <tr> <td>Rank</td> <td>Agent </td> <td>TrueSkill Rating</td> </tr> <tr> <td>1st</td> <td>Revac_8 </td> <td>13.9</td> </tr> <tr> <td>2nd</td> <td>Fractal_SecretMafia_Agent_round2_v25</td> <td>7.8</td> </tr> <tr> <td>3rd</td> <td>Fractal_SecretMafia_Agent_round2_v14u</td> <td>4.7</td> </tr> </tbody> </table> <h2><strong>Requirements</strong></h2> <ul> <li>Python 3.8+</li> <li>OpenAI API key or Groq API key</li> <li>Optional: Supabase credentials for game logging</li> </ul> <h2><strong>License</strong></h2> <p>MIT</p>