OASIS: Open Agent Social Interaction Simulations with One Million Agents

“To address these challenges, we employ a relationship network generation algorithm that combines a small amount of real user data to create a social network of up to one million users, while preserving the scale-free nature of social networks (Barab´asi & Albert, 1999).”

“Testing different models such as MiniLM v6 (Reimers & Gurevych, 2019), BERT (Devlin, 2018), and TwHIN-BERT.”

“In this section, we describe the workflow and critical internal mechanisms of OASIS , which enable it to be easily generalized and scaled to support the simulation of millions of LLM-based agents.”

“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.”

“Chain-of-Thought (CoT, Wei et al. (2022)) reasoning is incorporated,”

“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.”

“There have been numerous related studies, such as the exploration of multi-agent behavior patterns (Park et al., 2023), simulations of social networks (Gao et al., 2023; Zhou et al., 2023b), and the study of society’s response to misinformation (Chen & Shu, 2023).”

“Moreover, OASIS supports large-scale user interactions, accommodating up to 1 million users.”

“Recent simulator have drastically scaled these societies; for instance, frameworks like OASIS [29] and AgentSociety [30] achieve massive-scale simulations to study emergent social phenomena such as emotional contagion, and misinformation spread.”

“Frameworks such as AutoGen [37] and OASIS [39] provide general infrastructures for composing multiple LLM agents, coordinating their conversations, and supporting autonomous task completion or role-playing interaction.”

“We fix the scenario, agent personas, background context, and timeline size at 10 posts per timestep, and compare a follower-chronological baseline with two recommendation systems: a general embedding-based recommender and a TwHIN-based recommender [32], implemented following OASIS [11].”

“OA-SIS [213] builds a generalizable social-media sandbox with dynamic networks, recommendation mechanisms, rich action spaces, and million-scale agents to reproduce information spreading, group polarization, and herd effects.”

“…engines, social-media environments, learned or optimized surrogates, differentiable population frameworks, or topology-aware grouping [Gao et al., 2023, Tang et al., 2025, Yang et al., 2024, Prédhumeau and Manley, 2023, Jiang et al., 2024, Guan et al., 2025, Chopra, 2025, Xu et al., 2026].”

“Subsequent work has extended this paradigm to social-science experiments [32–36], user-behavior simulation [4, 5, 1–3, 37, 38], cognitive reasoning and human-response modeling [39–41], customer simulation [42, 43], marketing simulation [44–49], opinion dynamics and pluralism [50, 51], multi-turn…”

“Existing systems—including text-based simulations [25, 34, 29, 20] and human-platform studies [11]—remain limited to either text communication or human users.”

“In the domain of social media research, scholars have used LLM agents to study algorithmic polarization (Törnberg et al., 2023), misinformation spread (Yang et al., 2024), and the effects of content recommendation systems on user engagement and echo chamber formation (Chuang et al., 2023).”

“D AY (H O D) sampling (Yang et al., 2024), which constructs a per-agent activity histogram from historical data (e.g., over a 24-hour cycle).”

“February 9, 2026. be populated by powerful LLM agents (Dafoe et al., 2021; Chuang et al., 2024; Yang et al., 2024; Wu et al., 2024; Park et al., 2023), understanding their NE becomes important (Carichon et al., 2025; Yao et al., 2024).”

“For proactive assistants, LLM-driven simulators like GIDEA [47, 48] leverage this fidelity to generate “silicon samples,” allowing researchers to explore the vast space of user-assistant interactions that would be logistically impossible to observe in the real world [26].”

“…social interaction behaviors (Wei et al., 2022; Park et al., 2023; Zhao et al., 2024), multi-agent systems (MAS) composed of these models are widely used for human-like simulations (Park et al., 2023; Li et al., 2023; Yang et al., 2024; Li et al., 2024; Yang et al., 2025b; Wang et al., 2025).”

“Retrieval-style pipelines typically build personas from real-world data: some retrieve or enrich existing user personas for a given topic (Yang et al., 2024; Zhang et al., 2025), while others directly infer implicit personas from user logs (Wang et al., 2025a).”

“…& Corneil, 2025); to study user behavior in healthcare (Li et al., 2024), education (Gao et al., 2025), and social interaction (Wang et al., 2025; Yang et al., 2024; Tang et al., 2024); to help refine recommendation systems (Ma et al., 2025; Ebrat et al., 2024; Gu et al., 2025; Zhang et al.,…”

“…often concentrating activity among higher-capability contributors (Burtch et al. 2024, Shan and Qiu 2025), whereas AI-generated metadata (e.g., titles and hashtags) significantly boosts user engagement by improving user-content matching and inspiring better human-AI co-creation (Zhang et al. 2024).”

“…context; 2) Profile-Centric Models, PopAlign persona (Hu et al. 2025) and USP-style implicit profiling (Wang et al. 2025a) which use static persona grounding; 3) Social-Media Agents, The OASIS framework (Yang et al. 2024), representing current state-of-the-art in multi-agent social simulation.”

“Recent works also increase scale of LLM agent societies, with OASIS [5] achieving million-agent simulation for recommendation and social media phenomena, and AgentSociety [4] which simulates multiple scales of a society from mobility in a city to emergent social behaviors as polarization and…”

“…of LLM-empowered simulators generate agents with linguistic reasoning, spanning social-movement modelling 28 , million-user network environments 33,34 , opinion-dynamics exploration 35–37 , ethical-dilemma resolution 38 , malicious botnet behaviour 39 , graph growth 40 , recommender-system…”

“Recent work has explored using LLMs to emulate human behaviors (Zhou et al., 2024; Li et al., 2023; Xie et al., 2024; Yang et al., 2024).”

“To evaluate the dissemination effects, we employed 36 human-like agents simulated by OASIS (Yang et al., 2025b).”

“LLMs are applied in diverse multi-agent contexts: software development ( MetaGPT [35]), simulated scientific discovery [36, 37], complex social simulations [33, 3, 38, 39], and code generation [40].”

“ABMs have already been successfully applied to make predictions in economics (Axtell & Farmer, 2025; Vu et al., 2022), with LLM-based ABMs already simulating large-scale social interactions in digital environments such as Twitter and Reddit (Yang et al., 2024).”

“LLM-agent simulations study populations of thousands to millions of agents [12, 46, 47, 63], while benchmarks examine negotiation and communication failures [5, 10, 67].”

“Generative Agents models persistent individual routines and social interactions, while OASIS (Yang et al., 2024), AgentSociety (Piao et al., 2025), and Socio-Verse (Zhang et al., 2025) extend LLM-agent simulation to large-scale social environments.”

“…participation information : (1) Direct LLM ∗ , which generates actions directly from the event context [8, 46]; (2) Social Retrieval ∗ , which conditions generation on retrieved social traces [32, 41]; and (3) MF-LLM ∗ , which uses a recursively updated textual mean field to guide generation [26].”

“Yang et al. built OASIS, a large-scale Twitter and Reddit simulator capable of scaling to millions of agents, demonstrating information propagation and opinion herding dynamics [49].”

“We simulate a social media platform (adapted from Yang et al. (2024)) on which agents hold stateful accounts and can author posts and comments , including threaded replies.”

“We instantiate this reasoning as a single autoregressive trajectory, avoiding the cost and credit-assignment difficulties of a full multi-agent simulation (Yang et al., 2024).”

“Subsequent work scales such simulations to richer environments and larger populations (Yang et al., 2024, Piao et al., 2025), while recent research increasingly grounds simulated agents in real human data.”

“This vulnerability has important implications for personalized AI agents, which are becoming increasingly prevalent [41, 1, 83, 9].”

“Larger platforms followed, including OASIS with a million agents (Yang et al., 2024) and AgentSociety (Piao et al., 2025), alongside work that used simulated feeds to compare alternative ranking algorithms (Törnberg et al., 2023) and to trace opinion dynamics in networks of agents (Chuang et al.,…”

“The attacker knows the profile of the target user, for example, the user’s biog-raphy (Yang et al., 2024), which we use in the experiments.”

“…to an LLM would catastrophically II) for social, game theory and economic experimentation playgrounds such as CAMEL [115], Voyager [116], OASIS [117], EconAgent [118], AgentScope [119], [120], COMET [121], ASVO [123] and MoRE [122], our system integrates in either dynamically and/or…”

“However, these platforms produce collective behaviors and simulated populations that rely solely on computational, systematic design and the abilities of autonomous agents [8, 11–15].”

“Social media is a common setting for classical and LLM-based simulations (Yang et al., 2024; Lin et al., 2025; Butler et al., 2024; Jeon et al., 2025; Larooij and Törnberg, 2025; Cheng et al., 2025; Nguyen et al., 2026).”

“Constructing agents using real data also raises privacy concerns, while optimizing content reach may introduce risks related to persuasion and influence [15], [26], [28], [32].”

“…a thousand people’s survey answers (Park et al., 2024), urban simulators run policy experiments over tens of thousands of agents (Piao et al., 2025; Yang et al., 2024), and decentralized LLM populations appear to develop shared conventions and tipping points that mirror human collective dynamics…”

“While effective in relatively stable settings such as virtual societies and social networks (Park et al., 2023; Yang et al., 2024; Mou et al., 2024b; Tang et al., 2025), this paradigm fundamentally breaks down in socially intensive environments.”

“We compare CARD with OASIS and SynthPAI, two LLM-based frameworks that support social simulation, across four LLMs. OA-SIS [28] is a general and scalable social-media simulator that models user profiles, social networks, user actions, and recommendation mechanisms.”

“This paradigm includes "Generative Agents" [29], with studies focusing on individual-level persona modeling [14, 16, 31] and societal-level simulations of emergent behaviors [2, 21, 36].”

“OASIS [4] is a social-media simulation of up to a million agents, modelled on X and Reddit; it ships two recommender modes, and the mode is the perception switch.”

“Research on social bots and misinformation instead focuses on coordinated behavior in online ecosystems [11]–[13], and agent-based simulators model sequential user– platform interaction [14]–[18].”

“…agents on TikTok and YouTube that navigate content via scripted sequences; Bartley et al. [9] and Chen et al. [22] configure bots to perform simple interactions; other works replace scripted behavior with LLM agents, but deploy them in researcher-built simulated environments [25, 71, 78, 83].”

“Recent agent-based and LLM-driven simulations enrich this line by modeling adaptive, language-mediated agents to simulate broad social dynamics [47, 71].”

“This is important in OSN simulations, where user actions shape exposure, diffusion, and network structure [19, 7].”

“Agent-based frameworks demonstrate that large-scale social simulations are feasible [53, 54, 55], but computational costs remain prohibitive for systematic exploration.”

“Recent video diffusion models leverage the efficiency of latent space processing and the scalability and quality of transformer-based architectures.”

“Studies on personality-conditioned agents have further explored how traits influence cooperation in games (Qiu, 2025) and social media behavior in large-scale simulations (Yang et al., 2024).”

“Although simulation-driven approaches have demonstrated remarkable capabilities in modeling complex interactive settings [Park et al., 2023, Yang et al., 2024], their application to deliberation data generation remains constrained by several fundamental limitations.”

“Built on the OASIS framework (Yang et al. 2025), it constructs parallel digital worlds populated by thousands of LLM-driven agents with independent personality, memory, and behavioral logic, externalizing not individual cognition but the collective cognitive dynamics of social groups (opinion…”

“The implementation builds on the OASIS infrastructure (Yang et al. 2024) and introduces a modular, scalable architecture for controlled experimentation and counterfactual studies.”

“Agent-based simulations such as OASIS (Yang et al., 2024) model social media dynamics at massive scale but involve no human participants.”

“Recent advancements in MARL have explored more complex social dynamics, including cooperation (Guan et al., 2025), competition (Chen et al., 2024a), operation (Chen et al., 2024b), and the formation of social structures (Yang et al., 2024).”