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.”

“…2023] and role-playing of famous characters [Wang et al., 2023, Shao et al., 2023, Chen et al., 2024] or specific personalities [Caron and Srivastava, 2022], to society simulations [Argyle et al., 2023, Chuang et al., 2024, Yang et al., 2024, AL et al., 2024, Zhou et al., 2023, Pang et al., 2024].”

“However, the influence of the group on individuals often drives them to align their behavior with the group, leading to the emergence of the herd effect [33].”

“Recent studies have explored social simulation across various levels and scenarios, from mimicking well-known individuals and mirroring specific situations to modeling large-scale social dynamics [4, 29, 34, 36, 49, 60].”

“In our TwinMarket framework that leverages LLM-based agents to simulate complex human behaviors, we include rational decision-making, technical analysis, and behavioral biases such as herding and overconfidence [22, 23].”

“…of these studies include social media platform simulations (Gao et al., 2023; Papachristou and Yuan, 2024; Mou et al., 2024; Tang et al., 2024a; Yang et al., 2024), where LLM agents replicate human posting behaviors; financial markets (Yang et al., 2025), where they model investor trading…”

“Gao et al [23], Mou et al [86], Yang et al [87], Gu et al [90] and Ferraro et al [93] all based their validation on social media datasets on comparison with data collected from platforms such as Twitter or Reddit.”

“For social behaviors, we assess whether the agents merely engage in basic interactions [75] or demonstrate organized social relationships, reflecting more human-like group dynamics [109, 77, 6, 72].”

“Experiments show different topologies perform best for specific tasks [9, 53] and large-scale LLM agent networks exhibiting known social phenomena [48, 12].”

“Additionally, with the last generation of agents, we can evaluate long horizon tasks that are realistic and offer scalable testbeds for exploring coordination failures and emergent risks [Yang et al., 2024, Rossetti et al., 2024].”

“Notice that f(x) is monotonically increasing, so...”

“Some works [47, 74, 86] simulate sociological dynamics utilizing multiple agents, further expanding their capabilities.”

“We compare the performance of Llama3-8B-Instruct fine-tuned using MATRIX-Gen-Coding dataset against those SFT datasets in the coding domain or in the general domain, all in 10K data samples.”

“Recent advances in multi-agent simulations have substantially enhanced our ability to model the dynamics of individuals making decisions and interacting over time [1, 17, 21, 31, 35, 36, 40, 44, 57].”

“…models (LLMs) (Guo et al., 2025; Wei et al., 2025; Hu et al., 2025) can endow social agents with reasoning, perception, and interaction capabilities, enabling simulations from small-scale communities (Park et al., 2023; 2024) to multi-scene social systems (Wang et al., 2025; Zhang et al., 2025b).”

“These applications are expected to expand rapidly, as reflected in growing market interest, with valuations projected to increase from USD 5.4 billion in 2024 to USD 236 billion by 2034, and a surge in startups and enterprise features focused on interacting LLM agents (7, 8).”

“Earlier studies on sequential social dilemmas [36] and concurrent studies with our work on the limits of large-population models [9, 72] further underscore the role of heterogeneity and bounded rationality in shaping collective behavior.”

“They represent a groundbreaking technological advancement that enables direct analysis and computation of non-numerical data, including language and images.”

“Existing studies have applied LLM agents to domains such as economics [52, 69], education [148], game theory [117], and social networks [126, 140, 145], with claimed advantages like handling natural language, enabling flexible behaviors, and showing human-like reasoning.”

“With the advancement of role-playing abilities of large language models (LLMs), we can now simulate realistic social interactions at scale (Zhou et al., 2024c; Li et al., 2023; Pang et al., 2024; Yang et al., 2024).”

“State-of-the-art simulations such as OASIS [43] and AgentSociety [32] excel at environment modeling but remain prompt-centric in agent decision-making—relying on manually crafted role templates, heuristic memory-retrieval rules, and fixed interaction schedules.”

“Macro-level patterns emerge from micro-level agent behavior and meso-level interaction, as in social media discussion [70, 45], opinion dynamics and mobilization [37], polarization and echo chambers [19, 43, 62, 70], information or misinformation spread [31, 30, 40], conformity and herd effects…”

“where 𝜙 ( 𝑢 ) is the user profile and { 𝑚 } represents the user’s recent user commonly used evaluation metrics [46, 60, 67], the reward 𝑟 𝑡𝑖 𝑥 𝑡𝑖 Formally, given a post-user arm 𝑥 𝑡𝑖 and the reaction 𝑜 𝑡 + 1 𝑖 the reward for arm 𝑥 𝑡𝑖 where mis ( 𝑢 ) indicates whether user 𝑢 is…”

“More recent advances explore to automatically construct agentic environments from either pre-defined (Shi et al., 2025; Andrews et al., 2025; Yang et al., 2025a; Jain et al., 2025) or continuously refined (Hu et al., 2025c; Zhang et al., 2025b) tool and application suites.”

“Two designs could address this gap using logs that already exist (Zhao et al., 2024; Zheng et al., 2024; Kirk et al., 2024; Paruchuri et al., 2025) and telemetry MatrAIx already stores.”

“(see Table 1, Table 2, and Section 4).”

“For instance, many existing studies require dozens of GPUs and weeks of computation to simulate up to a million agents in social scenarios [29; 32; 28; 33; 36; 37; 38].”

“Subsequent frameworks such as SocioVerse [90], AgentSociety [61], Yulan-OneSim [80], and OA-SIS [87] further pushed this line toward more scalable infrastructures, richer agent coordination, and larger social environments, including simulations with up to one million agents.”

“Building on prior work [44], [52], we employ multiple collaborative LLM agents per employee, forming an “agent bundle” for greater behavioral realism.”

“This framework is built on OASIS, a social simulator that can support up to one million users [24].”

“Simulation platforms such as OASIS (Yang et al., 2024b), S 3 (Gao et al., 2023), and AgentSociety (Piao et al., 2025) study agent social dynamics at scales of up to one million agents, while Generative Agents (Park et al., 2023) demonstrates human-like social behavior in sandbox environments.”

“…under the guidance of high-level rules, resulting in organically cooperative behaviors and impressive performance in virtual tasks such as gaming (Wang et al., 2023a;b; 2024b; Qin et al., 2024), human simulation (Yang et al., 2024; Park et al., 2023), and disease diagnosis (Li et al., 2024a).”

“The other category is large-scale, LLM-driven multi-agent simulations (Yang et al. 2024b; Piao et al. 2025; Liu et al. 2025), designed to reproduce emergent behaviors in open-ended social settings and offer insights into collective dynamics.”

“…discussion: (i) it reduces mutual influence during initial assessment, allowing each expert to form an unbiased domain-specific opinion [35], (ii) it enables integration and update of previous insights through H t − 1 while avoiding false consensus [13], and (iii) it provides…”

“…prosperity of social media agents (Mou, Wei, and Huang 2024; Zhang et al. 2025a), which can potentially take on various tasks such as social event analysis (Qian, Zhang, and Xu 2016), post and user recommendation (Tang, Tang, and Liu 2014), and social behavior simulation (Yang et al. 2024).”

“0 [2] leverages AI-powered traffic flow simulation with reinforcement learning and synthetic data to establish high-27381 fidelity closed-loop environments, rigorously validating self-driving systems’ rare scenario handling capabilities.”

“Similarly, [24] takes this further by introducing OASIS, a scalable and extensible social media simulator.”

“…agent-based modeling to derive emergent macro-patterns from individual micro-decisions (Gao et al., 2023; Zhang et al., 2025b; Lorig et al., 2021; Piao et al., 2025; Yang et al., 2024), often emphasizing theory-of-mind modeling (Zhou et al., 2025a) or resource distribution (Zhang et al., 2025a).”

“Several benchmarks broaden evaluation: MultiTrust (Zhang et al., 2024) measures truthfulness, robustness, safety, fairness, and privacy, while other work emphasizes transparency and bias reduction in multi-modal systems (Saleh and Tabatabaei, 2025).”

“Another deploys LLM agents in social environments, from social-network simulations [50] to agent societies [51, 52], modeling phenomena such as information diffusion [53] and negotiation [54].”

“Second, it induces fragility and hidden confounding: because behavior is prompt-and interface-mediated, small changes in persona wording, memory format, tools, or interaction protocol can shift conclusions later interpreted as substantive policy or social-mechanism effects [49, 33, 80].”

“As a result, these models are being adopted as “promising” [18] tools in computational social science to create human-like agents [26, 27] and to analyze multi-agent interactions in complex social systems [28, 29, 30, 31, 32].”

“Recent LLM research leverages role-play to simulate individuals with different cultural backgrounds for social simulations [172–174] with population-level scalability [175].”

“Scalable Simulation To better replicate the phenomenon of free collaboration in real scientific cooperation, we implement an adaptive concurrent distributed system based on the OASIS [23].”

“State-of-the-art simulations such as OASIS [43] and AgentSociety [32] excel at environment modeling but remain prompt-centric in agent decision-making—relying on manually crafted role templates, heuristic memory-retrieval rules, and fixed interaction schedules.”

“…imagine a complex multi-agent system made of models trained on scientific texts and assigned synthetic deliberative personas (Chapuis et al. 2022; Yang et al. 2025), capable of engaging with abstract and concrete research questions alike, generating solutions and objections in indefinite…”

“Each simulation ran for 90 minutes, a duration aligned with recent work in the domain of synthetic OSNs simulations (e.g., [9]) and su ffi cient to allow meaningful interaction dynamics, content propagation, and behavioral di ff erentiation to emerge.”

“Recent work on large–scale social simulators like OASIS demonstrates that dynamic social graphs coupled with recommendation mechanisms reproduce realistic adoption patterns at scale [48].”

“…extends this interactive view to role-playing, coordination, and social simulation, showing that agent behavior depends on role structure, protocol, and the surrounding social setting (Park et al., 2023; Li et al., 2023; Wu et al., 2024; Gao et al., 2023; Piao et al., 2025; Yang et al., 2024).”

“…work has developed richer persona representations (Zhang et al., 2018; Ge et al., 2024; NVIDIA, 2025; Wang et al., 2025b), while LLM-based agent frameworks make it easier to deploy these personas as interactive agents (Park et al., 2023; Vezhnevets et al., 2023; Yang et al., 2024; Microsoft, 2025).”

“…simulation studies examine multi-agent interaction and social behavior in modeled societies (Park et al. 2023; Piao et al. 2025; Jiang et al. 2024; Yang et al. 2024; Song et al. 2025), while social-navigation, human–robot interaction, privacy-aware, and interactive-safety evaluations focus on…”

“The framework supports multiple LLM backends, including OpenAI and Anthropic APIs as well as self-hosted models via SGLang (Zheng et al., 2024), allowing experiments to be run with both proprietary and open-source models.”

“Early work showed that agents equipped with memory, planning, and communication capabilities can coordinate activities, form social relationships, and exhibit emergent collective behavior [14,23,66,93,100].”

“These population sizes reflect document healthcare IT environments, covering the staffing levels of small community hospitals up through large academic medical centers and multi-site health systems [58,59].”

“They can also each reach the two corner bins on either side of the area in front of them, totaling four destination bins.”

“Some simulators escalate the number of agents to over 100,000 by activating a fixed subset of core agents to engage in human-like interactions with a static hierarchical structure [14], [18], [21].”

“Static topologies define fixed communication graphs, as in systems such as MEDCO [37] and Welfare Diplomacy [48], while dynamic topologies adapt in real time using models such as DyLAN [49], GPTSwarm [50], and OASIS [61].”

“Generative Agents (Park et al., 2023) demonstrated that memory and reflection can support believable social behaviour, while subsequent systems,including S 3 (Gao et al., 2023), AgentSociety (Piao et al., 2025), and OASIS (Yang et al., 2024), have extended LLM-based social simulation to larger populations.”

“Recent social simulators such as OASIS [120] scale to one million agents, revealing diverse social dynamics, including group polarization and herd behavior.”

“More recent work has pushed this paradigm toward larger-scale and more empirically grounded simulation, including efforts to simulate populations and known human-group dynamics at scale [7–9, 29, 35].”

“The studies are [122], [123], [124], [125], [126], [127], [128].”

“Other GABM frameworks focus only on interaction on social media, like OASIS 51 , capturing verbal interaction within social and algorithmic systems.”

“The categories of RPLAs span from fiction characters and celebrities to daily life characters, supporting various scenarios, e.g. , sociological studies(Yang et al., 2024; Huang et al., 2024; Tang et al., 2025; Wang et al., 2024b).”

“To ensure actions are executable, an Action Simulator runs pre-execution rollouts to detect constraint violations and triggers local repair; execution failures invoke a tiered re-planning pipeline from fast Short-Term Memory based fixes to full top-down replanning when necessary.”

“Other systems such as OASIS [40 ] and SocioVerse [31 ] focus on large-scale simulations, incorporating dynamic contexts, recommendation mechanisms, and reconstructed real-world information to model polarization and misinformation cascades.”

“…Radford et al. (2018); Kojima et al. (2022) can simulate human emotions and opinions, from subjective labeling of Twitter posts (Törnberg, 2023; Yang et al., 2024) and participation in psychological studies (Aher et al., 2023; Qiu & Lan, 2024) to producing behavior changes consistent with…”

“Recent studies have leveraged generative agents to simulate complex dynamics within social networks, aiming to explore emergent human-like behaviors [16, 47, 52, 8].”

“…resembles the world: whether agents behave plausibly (Park et al., 2023), whether outcomes match empirical patterns (Larooij & T¨ornberg, 2026; Yang et al., 2024; Gu et al., 2025 Cooperation is prevalent in repeated social dilemmas. supportable from PLAIN runs Defection is prevalent in…”

“…on believable generative agents, researchers have rapidly adopted this approach across a range of social simulation studies (Ferraro et al., 2024; Yang et al., 2024; Gao et al., 2023; T¨ornberg et al., 2023; Composta et al., 2025; Larooij & T¨ornberg, 2025a; De Marzo et al., 2023; Bouleimen et…”

“Subsequent systems study opinion formation, diffusion, polarization, and platform-scale interaction (Chuang et al., 2024; Yang et al., 2024; Zhang et al., 2025).”

“In the reported implementation, OASIS provides accounts, relations, posts, comments, likes, feeds, traces, and platform-compatible timestamps as part of this environment [4].”

“For example, OASIS [39] initializes agents with persona prompts derived from demographic attributes to support large-scale social simulations, while Persona Hub [12] curates extensive per-sona libraries to induce variation in synthetic data generation.”

“Recent systems have scaled agent­based simulations to the order of 10 4 LLM agents using distributed orches­ tration and efficient memory management 60,63,64 .”