Among Us Social Deduction Game

该数据集模拟了游戏《Among Us》中的多代理环境，其中包含了代理之间的互动与沟通数据。这些数据旨在让研究者分析在社交推理游戏中，如识别伪装者的过程中，沟通策略和涌现行为。此外，该数据集还关注了胜率以及沟通的有效性。该数据集已用于评估拥有1.5亿和70亿参数的模型，任务集中在社交推理游戏中的多代理强化学习。

This dataset simulates the multi-agent environment within the game *Among Us*, and contains interaction and communication data between agents. It is designed to enable researchers to analyze communication strategies and emergent behaviors during social deduction gameplay, such as the process of identifying impostors. Additionally, this dataset also focuses on win rates and the effectiveness of communication. This dataset has been used to evaluate models with 150 million and 7 billion parameters, with tasks focused on multi-agent reinforcement learning in social deduction games.

搜集汇总

背景与挑战

背景概述

该数据集围绕《Among Us》社交推理游戏，研究如何通过多智能体强化学习训练语言模型进行自然语言交流，以提高智能体的社交推理能力。研究通过分解交流为‘听’和‘说’两部分，并利用环境信息预测作为奖励信号，实现了在不依赖人类演示的情况下训练语言模型，显著提高了游戏中的胜率。