OASIS: Open Agent Social Interaction Simulations with One Million Agents

OASIS：拥有百万代理的开放代理社交交互模拟

1 Experiment 实验

Although OASIS has the potential to be applied for various computational inquiries, we primarily focus on two research questions below:
尽管 OASIS 具有应用于各种计算查询的潜力，我们主要关注以下两个研究问题：

  1. Can OASIS be adapted to various platforms and scenarios to replicate real-world phenomena? We demonstrate the generalizability of OASIS by replicating three influential computational social science studies. Specifically, we simulate information propagation (Vosoughi et al., 2018) and the resulting group polarization (Lindesmith et al., 1999) on rapid information exchange platforms like X and the herd effect (Muchnik et al., 2013) on topic-based community-oriented platforms like Reddit.
    OASIS 能否适应各种平台和场景以复制现实世界现象？我们通过复制三个有影响力的计算社会科学研究来展示 OASIS 的泛化能力。具体来说，我们在像 X 这样的快速信息交换平台上模拟信息传播（Vosoughi 等人，2018 年）及其导致的群体极化（Lindesmith 等人，1999 年），以及在像 Reddit 这样的基于主题的社区导向平台上模拟羊群效应（Muchnik 等人，2013 年）。
  2. Does the agent population affect the accuracy of simulating group behavior? We conduct sociological experiments at various scales of agents, ranging from hundreds to tens of thousands of agents, and identify (if any) emergent sociological phenomena as the number of agents increases.
    代理人口数量是否会影响模拟群体行为的准确性？我们在不同规模的代理人口中进行社会学实验，从数百到数万不等，并随着代理人口数量的增加，识别（如果有的话）出现的（社会）现象。

1.1 Experimental Scenarios 实验场景

Information propagation on X.
信息在 X 上的传播
Information propagation refers to the propagation of messages through a network, influenced by varied factors (e.g., network structure, message content, and individual interactions). It is crucial for understanding phenomena like information spreading and group polarization. In this section, we explore two key aspects: information spreading, the transmission of messages across a network; and group polarization, where social interactions foster increasingly extreme opinions. Our analysis focuses on these dynamics within the X platform.
信息传播指的是消息通过网络传播，受多种因素影响（例如，网络结构、消息内容和个体互动）。了解信息传播和群体极化等现象至关重要。在本节中，我们探讨两个关键方面：信息传播，即消息在网络中的传输；以及群体极化，即社会互动导致越来越极端的观点。我们的分析集中在 X 平台上的这些动态。

Herd effect in Reddit.
羊群效应在 Reddit 上。
Herd effect refers to individuals’ tendency to follow the actions or opinions of a larger group without independent thought or analysis. For example, users tend to like a post that has already received likes or reflect a general inclination to conform to majority opinions. Our analysis focuses on these dynamics within the Reddit platform.
羊群效应指的是个人在没有独立思考或分析的情况下，倾向于跟随更大群体的行为或观点。例如，用户倾向于喜欢已经获得点赞的帖子，或者表现出一种倾向于遵守多数人观点的倾向。我们的分析集中在 Reddit 平台上的这些动态。

3.2Experimental Settings 3.2 实验设置

For information spreading, we collect 198 real-world instances from two rumor detection datasets, Twitter15 (Liu et al., 2015) and Twitter16 (Ma et al., 2016), covering 9 categories (e.g., business, education, and politics). Each instance includes 100 to 700 users and the information propagation path of the source post. Using the X API, we retrieve user profiles, follow relationships, and previous posts, computing users’ hourly activity levels (See Appendix E.1 for details). Agents in OASIS are initialized with this data, and their most recent posts will also be included in the simulator to be propagated along with the source post for better alignment with real-world scenario (Section 2.1). For group polarization, we select 196 real users’ information from the information-spreading experiment (these real users have a large following on X and they are from different areas.) and using LLMs to generate synthetic users with up to 1 million scale (Prompts and details are presented in Appendix E.2). Real users are set as core users, with generated users forming follow-up relationships based on topics like sports and entertainment. For herd effect, we first closely follow Muchnik et al. (2013) and collect 116,932 real comments from Reddit across seven topics and use LLMs to generate profiles for 3,600 users. Second, we collect 21,919 counterfactual content posts (Meng et al., 2022) and generate 10,000 users. Comments or posts are divided into three groups: the down-treated group (one initial dislike), the control group (no initial likes or dislikes), and the up-treated group (one initial like). We simulate 40 or 30 time steps of interactions for each experiment on Reddit, introducing initially-rated comments or posts at the beginning of each time step (Details are presented in Appendix E.3 and F.4.2). Llama3-8b-instruct is used as the base LLM. We adjust agent actions to accommodate different scenarios, with specific actions for each scenario detailed in Appendix F.1.
为了信息传播，我们从两个谣言检测数据集 Twitter15（刘等，2015）和 Twitter16（马等，2016）中收集了 198 个真实世界实例，涵盖 9 个类别（例如，商业、教育和政治）。每个实例包括 100 到 700 个用户和源帖子的信息传播路径。使用 X API，我们检索用户资料、关注关系和以前发布的帖子，计算用户的每小时活动水平（详细信息请见附录 E.1）。OASIS 中的代理使用这些数据初始化，并且他们的最新帖子也将被包含在模拟器中，与源帖子一起传播，以更好地与真实世界场景对齐（第 2.1 节）。对于群体极化，我们从信息传播实验中选择了 196 个真实用户的信息（这些真实用户在 X 上拥有大量关注者，并且来自不同的地区），使用LLMs生成最多达到 100 万规模的合成用户（提示和详细信息请见附录 E.2）。真实用户被设置为核心用户，生成的用户根据体育和娱乐等主题形成后续关注关系。对于羊群效应，我们首先紧密遵循 Muchnik 等人的方法。 (2013) 并从七个主题中收集了 116,932 条 Reddit 真实评论，使用LLMs为 3,600 个用户生成档案。其次，我们收集了 21,919 篇反事实内容帖子（Meng 等人，2022 年）并生成了 10,000 个用户。评论或帖子分为三组：下行处理组（一个初始不喜欢），对照组（没有初始喜欢或不喜欢），和上行处理组（一个初始喜欢）。我们对 Reddit 上的每个实验模拟了 40 或 30 个时间步长的交互，在每个时间步长的开始引入初始评分的评论或帖子（详细信息见附录 E.3 和 F.4.2）。Llama3-8b-instruct 用作基础LLM。我们调整代理行为以适应不同场景，每个场景的具体行为在附录 F.1 中详细说明。