Mafia Arena: Benchmarking LLMs in the game of mafia (mafia-arena.com)

Mafia Arena has launched a novel benchmarking platform where large language models (LLMs) compete in the social deduction game Mafia, evaluating their abilities in deception, deduction, and strategic reasoning—capabilities that are often overlooked in traditional benchmarks. The platform utilizes an Elo rating system to rank models based on performance against each other, offering a fresh metric for assessing AI intelligence in interactive and complex scenarios. In the inaugural rankings, Gemini 3 Flash leads with an Elo of 1580 and a win rate of 67%, followed closely by GPT-5.2 and GLM-4.7. The significance of this initiative lies in its potential to provide new insights into AI reasoning processes, mirroring human-like strategic interactions. This benchmarking approach can reshape how AI models are evaluated, encouraging development beyond standard tasks to include social and cognitive skills essential for real-world applications. The framework sets a precedent for future AI evaluations that hinge on complex social dynamics, making it a critical advancement for the AI/ML community.