PaperlayerView exact revision on arXiv

Source-linked AI summary

OASIS: Open Agent Social Interaction Simulations with One Million Agents

Ziyi Yang, Zaibin Zhang, Zirui Zheng, Yuxian Jiang, Ziyue Gan, Zhiyu Wang, Zijian Ling, Jinsong Chen, Martz Ma, Bowen Dong, Prateek Gupta, Shuyue Hu, Zhenfei Yin, Guohao Li, Xu Jia, Lijun Wang, Bernard Ghanem, Huchuan Lu, Chaochao Lu, Wanli Ouyang, Yu Qiao, Philip Torr, Jing Shao

arXiv:2411.11581v5cs.CL

TL;DR

Existing LLM-based social-media ABMs are difficult to generalize across platforms and typically simulate far fewer agents than real platforms contain. OASIS provides a modular, scalable simulator with dynamic environments, diverse actions, recommendation systems, and support for up to one million users. It reproduces information spreading, polarization, and herd effects across X and Reddit, while showing stronger group dynamics and more diverse, helpful opinions at larger scales.

  • Problem

    Existing LLM-based ABMs are often scenario-specific and limited in scale, despite social-media platforms having diverse interaction designs and millions of users.

    Source: §1 · p. 2

  • Method

    OASIS uses modular social-media components, recommendation systems, temporal activation, and scalable inference to simulate up to one million users.

    Source: §1 · p. 3, §2.5 · p. 5, §5 · p. 12

  • Results

    OASIS replicates information spreading, group polarization, and herd effects across X and Reddit, with larger groups producing stronger dynamics and more diverse, helpful opinions.

    Source: §1 · p. 3

  • Takeaways & Limitations

    OASIS provides a generalizable environment for studying large-scale social group dynamics and multi-agent interactions.

    Source: §1 · p. 3, §5 · p. 12

    The simplified recommendation system cannot accurately represent nuanced intermediary-user profiles because of data limitations.

    Source: §3.3.1 · p. 8

Abstract

from arXiv · show

There has been a growing interest in enhancing rule-based agent-based models (ABMs) for social media platforms (i.e., X, Reddit) with more realistic large language model (LLM) agents, thereby allowing for a more nuanced study of complex systems. As a result, several LLM-based ABMs have been proposed in the past year. While they hold promise, each simulator is specifically designed to study a particular scenario, making it time-consuming and resource-intensive to explore other phenomena using the same ABM. Additionally, these models simulate only a limited number of agents, whereas real-world social media platforms involve millions of users. To this end, we propose OASIS, a generalizable and scalable social media simulator. OASIS is designed based on real-world social media platforms, incorporating dynamically updated environments (i.e., dynamic social networks and post information), diverse action spaces (i.e., following, commenting), and recommendation systems (i.e., interest-based and hot-score-based). Additionally, OASIS supports large-scale user simulations, capable of modeling up to one million users. With these features, OASIS can be easily extended to different social media platforms to study large-scale group phenomena and behaviors. We replicate various social phenomena, including information spreading, group polarization, and herd effects across X and Reddit platforms. Moreover, we provide observations of social phenomena at different agent group scales. We observe that the larger agent group scale leads to more enhanced group dynamics and more diverse and helpful agents' opinions. These findings demonstrate OASIS's potential as a powerful tool for studying complex systems in digital environments.

1 INTRODUCTION

OASIS addresses the limited generalizability and scale of existing LLM-based social-media ABMs by providing a modular simulator for diverse platforms and phenomena. It supports large-scale users and reproduces several social dynamics across X and Reddit.

Motivation

  • Existing LLM-based ABMs are often scenario-specific, while social-media platforms differ in action spaces, information filters, and dynamic networks.These differences make adapting simulators to other domains tedious, while real platforms contain millions of users.Source: §1 · p. 2

Contribution

  • OASIS combines adaptable environment, filtering, agent, timing, and inference components that can be modified across platforms such as X and Reddit.Its modular design supports platform-specific interactions and recommendation behavior.Source: §1 · p. 2

Scalability

  • OASIS supports simulations ranging from hundreds to millions of agents, addressing the under-explored importance of scale in LLM-based ABMs.The simulator includes user generation and multiprocessing techniques for large-scale experiments.Source: §1 · pp. 2–3

Scale effects

  • Larger agent groups produce stronger group dynamics and more diverse and helpful agent opinions.The findings connect simulation scale with the quality and diversity of observed group behavior.Source: §1 · p. 3

Empirical validation

  • OASIS reproduces information spreading, group polarization, and herd effects across X and Reddit, while also revealing distinct behaviors in LLM-driven societies.Uncensored LLMs show more severe polarization, and agents are more susceptible to herd effects than humans.Source: §1 · p. 3

2 METHODOLOGY

OASIS is a generalizable, scalable LLM-based social-media simulator built from modular platform components, dynamic environments, recommendation systems, diverse agent actions, temporal behavior, and distributed inference. Its methodology supports simulations ranging from individual interactions to one million users across X and Reddit.

2 METHODOLOGY

  • OASIS combines an Environment Server, RecSys, Agent Module, Time Engine, and Scalable Inferencer to support generalizable and million-agent social-media simulation.The architecture is designed to model social-media platforms while enabling large-scale LLM-agent inference.Source: §2 · p. 3, §2.1 · p. 3, §2.6 · p. 6

2.1 WORKFLOW OF OASIS

  • Users register profiles and histories, then receive character and action descriptions that guide platform-specific behavior during simulation.The environment recommends contextual posts, after which agents select actions and produce reasoning before updating the environment.Source: §2.1 · p. 3, Figure 2 · p. 4

2.2 ENVIRONMENT SERVER

  • The Environment Server stores users, posts, comments, relationships, recommendations, and complete action traces in a relational database that updates dynamically.New users, posts, comments, and follow relationships can be added over time.Source: §2.2 · p. 4

2.3 RECSYS

  • The RecSys controls agents’ information exposure using popularity-ranked in-network posts, interest-matched and recency-adjusted out-of-network posts on X, and hot-score ranking on Reddit.The system supports adjustable in-network and out-of-network recommendation counts; Reddit’s hot score integrates votes and submission time.Source: §2.3 · pp. 4–5, Figure 3 · p. 5

2.4 AGENT MODULE

  • The Agent Module combines memory and action components, supports 21 interaction types, and uses chain-of-thought reasoning to improve behavioral interpretability and interaction diversity.Memory includes post engagement information, prior actions, and the reasoning behind those actions.Source: §2.4 · p. 5

2.5 TIME ENGINE

  • Agents are activated probabilistically from 24-dimensional hourly activity profiles rather than simultaneously, while three-minute time steps accommodate different inference speeds.An alternative time-flow setting preserves within-step ordering for Reddit post timestamps.Source: §2.5 · pp. 5–6

2.6 SCALABLE DESIGN

  • Scalable design separates agents, the environment, and inference services into concurrent modules, while GPU management balances requests across available resources.The user-generation algorithm combines real data and network modeling to simulate up to one million users while preserving scale-free network structure.Source: §2.6 · p. 6

3 EXPERIMENT

OASIS evaluates whether a scalable simulator can reproduce social phenomena across X and Reddit and how agent population size affects simulated group behavior. Across experiments, it matches several real-world patterns while revealing scale-dependent opinion diversity, helpfulness, herd effects, and misinformation dynamics.

3 EXPERIMENT

  • OASIS evaluates generalizability across platforms by simulating information propagation and group polarization on X and herd effects on Reddit, while varying agent population size.The experiments address platform adaptation and whether population scale affects group-behavior simulation.Source: §3 · p. 6

3.3.1 INFORMATION PROPAGATION IN X

  • OASIS reproduces X information spreading in scale and maximum breadth with about 30% normalized RMSE, but its propagation depth remains below real-world trends.The comparison covers 198 real-world propagation instances; the authors attribute the depth discrepancy partly to simplified recommendation systems and user profiles.Source: §3.3.1 · pp. 7–8, §3.2 · p. 7, Figure 4 · p. 8
  • During X information propagation, agent opinions become increasingly conservative and extreme, with stronger polarization in uncensored models.The uncensored model more often produces extreme phrasing as interactions progress.Source: §3.3.1 · p. 8, Figure 5 · p. 8

3.3.2 HERD EFFECT IN REDDIT

  • On Reddit, agents show stronger herd behavior than humans for down-treated comments, while their responses to up-treated comments are relatively similar.Agents tend to further dislike or give fewer likes after an initial dislike, whereas humans more often deliberate and increase the like score.Source: §3.3.2 · p. 9, Figure 6 · p. 9

3.4.1 INFORMATION PROPAGATION IN X

  • Increasing the X agent population from 196 to 10,196 and 100,196 produces more diverse opinions and progressively higher helpfulness for core users.The authors connect this pattern to core users receiving a broader and more enriching set of responses.Source: §3.4.1 · p. 9, Figure 7 · p. 10

3.4.2 HERD EFFECT IN REDDIT

  • For counterfactual Reddit posts, herd effects emerge only for dislikes and become more pronounced as the population grows from 100 to 10,000 agents.The down-treated group develops significantly higher disagree scores than the control and up-treated groups, alongside increased scores interpreted as self-correction.Source: §3.4.2 · pp. 9–10, Figure 8 · p. 10

3.5 SIMULATING LARGE-SCALE MISINFORMATION SPREADING ON PLATFORM X USING OASIS

  • In a one-million-agent X simulation, misinformation generates more related posts than official news across all four topics and retains greater activity later in the simulation.The experiment used 196 core agents, one million regular agents, and 733,824 simulated posts analyzed with TF-IDF cosine similarity.Source: §3.5 · pp. 10–11, Figure 9 · p. 11
  • New follow relationships formed during the misinformation simulation cluster into dense central areas, isolated users, and distinct community-like subgraphs.The visualized connections indicate structured relational networks emerging during simulation.Source: §3.5 · p. 11, Figure 10 · p. 11

4 ABLATION STUDY

The ablations show that both recommendation and real-world temporal activity patterns are important for reproducing information dissemination.

4.1 ABLATION OF COMPONENTS IN OASIS

  • Removing the RecSys significantly hampers information spread, while replacing real-world activity probabilities with uniform values reduces fidelity to observed dissemination patterns.The primary metric is Normalized RMSE measured every minute.Source: §4.1 · p. 12
  • TwHIN-BERT performs particularly well for capturing similarities between posts in the recommendation ablation.Source: §4.1 · p. 12
  • The temporal feature uses a 24-dimensional activity probability list extracted from users’ previous posting frequencies.Source: §4.1 · p. 12

5 CONCLUSION

OASIS is presented as a generalizable, scalable social-media simulator for studying complex social dynamics with large populations of LLM-based agents. Its motivation arises from the risks and complexity of social media, the limitations of traditional and small-scale simulations, and the need to study collective behavior at scale.

5 CONCLUSION

  • OASIS models real-world social media dynamics, supports up to 1 million users, reproduces known phenomena, and reveals group behaviors that vary with population size.Source: §5 · p. 12

B.1 SOCIAL MEDIA

  • Social media’s complex networks, vast data, diverse behaviors, and ethical concerns motivate controllable virtual environments for testing social hypotheses.Source: §B.1 · p. 19

B.2 MULTI-AGENT SYSTEMS

  • Multi-agent systems support differentiated roles and richer interactions, including collaboration, discussion, and strategic competition.Source: §B.2 · p. 20
  • Existing multi-agent systems include small tool-based assistants and larger society or game simulations, with capability scalability central to large social environments.Source: §B.2 · p. 20

B.3 MULTI-AGENT SYSTEM SOCIAL SIMULATION

  • LLM-based social simulations improve natural-language interaction and behavioral realism, but collective phenomena often require a critical mass of agents.Source: §B.3 · p. 20

C.1 MORE EFFICIENCY ANALYSIS

The efficiency analysis concerns the counterfactual herd-effect experiment and compares efficiency across different agent scales.

C.1 MORE EFFICIENCY ANALYSIS

  • Table 2 analyzes experiment efficiency across different agent scales for the counterfactual herd-effect experiment in Reddit.Source: §C.1 · p. 20, Table 2 · p. 20
  • The reported analysis focuses on efficiency in a Reddit experiment rather than on social-phenomenon outcomes.Source: §C.1 · p. 20
  • Table 2 is specifically titled as an efficiency analysis of different agent scales.Source: Table 2 · p. 20

C.2 RECOMMEND SYSTEM ABLATION

The recommendation-system ablations evaluate both removing the RecSys and changing its embedding model. They show that recommendation sustains dissemination and that TwHIN-BERT produces stronger recommendation behavior than MiniLM.

C.2 RECOMMEND SYSTEM ABLATION

  • The ablation study examines both RecSys existence and the models used to embed posts and profiles across selected propagation topics.Source: §C.2 · pp. 20–21
  • Removing the RecSys usually ends information spread prematurely, often reducing dissemination to broadcasts from a single superuser.Source: §C.2 · p. 21
  • The RecSys is essential for connecting isolated nodes and sustaining simulations, although dense follower networks can make its removal effective for some entertainment topics.Source: §C.2 · p. 21
  • TwHIN-BERT and regular BERT outperform paraphrase-MiniLM-L6-v2 in Figure 11a.Source: §C.2 · p. 21, Figure 11 · p. 21
  • TwHIN-BERT recommends more appropriate posts than the comparison models in the recommendation results shown in Figure 11b.Source: §C.2 · p. 21, Figure 11 · p. 21

C.3 TEMPORAL FEATURE ABLATION

Removing OASIS's temporal feature weakens its ability to reproduce real-world information-propagation dynamics. The ablation compares normalized RMSE across 28 topics with all agents assigned an activity probability of 1.0.

C.3 TEMPORAL FEATURE ABLATION

  • Figure 12 compares normalized RMSE for OASIS, OASIS without temporal-feature simulation, and real propagation.Source: Figure 12 · p. 21, §C.3 · p. 22
  • Without temporal features, OASIS cannot capture real-world information-propagation dynamics well because agents act too frequently.The ablation sets all activity probabilities to 1.0 and evaluates normalized RMSE on 28 topics.Source: §C.3 · p. 22

C.4 LLM ABLATION

The LLM ablation evaluates several open-source language models as agent backends for reproducing real-world information propagation. The accompanying action-prompt specification defines the available user actions and structured response format.

C.4 LLM ABLATION

  • Qwen1.5-7B-Chat, Internlm2-chat-20b, and Llama-3-8B-Instruct are tested as agent backends on 28 information-propagation topics.Source: §C.4 · p. 22
  • Figure 13 reports normalized RMSE for simulation results produced by different LLM-based agents.Source: Figure 13 · p. 22, §C.4 · p. 22

D.1 USER ACTIONS PROMPTS

  • The action space includes posting, reposting, liking, commenting, following, searching, refreshing, trending, muting, and doing nothing.Source: §D.1 · pp. 22–23
  • Agents are prompted to keep actions consistent with their self-description and return function names with arguments in directly convertible JSON.Source: §D.1 · p. 24

D.2 ENVIRONMENT SERVER DATABASE STRUCTURE

The environment server stores platform interactions in dedicated database tables and delivers recommendation-selected posts to agents. The system also supports asynchronous inference communication and GPU-aware request distribution.

D.2 ENVIRONMENT SERVER DATABASE STRUCTURE

D.3 RECOMMENDATION SYSTEM

  • The recommendation system ranks posts, stores top-ranked entries in a shared recommendation table, and samples linked post IDs when an agent refreshes its feed.Source: §D.3 · p. 25
  • For X, recommendation scoring uses recency, fan count, and cosine similarity between post and user-content embeddings.Source: §D.3 · p. 26

D.4 PARALLEL OPTIMIZATION

  • The information channel uses an asynchronous message queue and UUID-keyed response storage to coordinate concurrent agent requests.Source: §D.4 · p. 26, Figure 14 · p. 26
  • The inference manager distributes requests across GPU devices and can scale the number of graphics cards up or down.Source: §D.4 · p. 26

E.1 REAL-WORLD PROPAGATION DATA

This section describes the data sources, scenario-specific actions, propagation metrics, reproducibility setup, and evaluation protocols used in OASIS. It also details a counterfactual-content herd-effect experiment varying agent population size and simulation duration.

E.1 REAL-WORLD PROPAGATION DATA

  • OASIS samples 198 real propagations, enriches users with profiles, networks, previous posts, GPT-3.5 Turbo character descriptions, and hourly activity probabilities.Previous posts span three days before each source post, with selected time windows and prior-hour posts included as noise; activity probability Pij equals fij divided by maxk(fkj).Source: §E.1 · p. 27

E.2 GROUP POLARIZATION

  • The user-generation algorithm combines limited real data with demographic sampling and interest-based links to construct scale-free social networks of up to one million users.Profiles sample age, MBTI, professions, and social-network trends, while agents follow core users with aligned interests at probability 0.1.Source: §E.2 · pp. 27–28

E.3 HERD EFFECT

  • Reddit users are generated from demographic distributions, GPT-3.5 Turbo-selected interests across seven categories, and generated names, bios, and personas.The herd-effect datasets include 116,932 comments from 17 subreddits and 21,919 counterfactual entries, with generated posts aligned to the seven topics.Source: §E.3 · pp. 28, 30

F.1 ACTIONS OF DIFFERENT SCENARIOS

  • Scenario-specific action spaces are integrated into OASIS so users can freely select and combine actions across experiments.The available actions differ by scenario and platform, as summarized in Table 16.Source: §F.1 · p. 31, Table 16 · p. 31

F.2.1 METRICS

  • Propagation is evaluated by scale, depth, maximum breadth, and normalized RMSE against real-world propagation trajectories.Scale counts unique users, depth is the greatest distance from the root, and maximum breadth is the largest node count at any depth.Source: §F.2.1 · p. 31

F.2.2 ALIGN WITH REAL PROPAGATIONS

  • Propagation simulations run for 50 three-minute steps using only like, repost, follow, and do nothing, comparing 150 simulated minutes with real data.A 300-agent simulation for 30 time steps takes 26 minutes on one NVIDIA A100-SXM4-80GB; repeated trials remain relatively stable, though added posts and prominent users can disturb results.Source: §F.2.2 · p. 32, Figure 15 · p. 32

F.3.1 DILEMMA QUESTIONS / F.3.2 POLARIZATION EVALUATION PROMPTS / F.3.3 HELPFULLNESS EVALUATION PROMPTS

  • Polarization and helpfulness are assessed with dilemma scenarios, pairwise extreme-tone judgments, and comparisons of response quality across accuracy, clarity, completeness, context, creativity, depth, professionalism, sources, and engagement.The polarization prompt returns Answer1, Answer2, or neutral, while the helpfulness prompt compares two AI assistants.Source: §F.3.1 · p. 33, §F.3.2 · p. 33, §F.3.3 · pp. 33–34

F.4.1 METRICS / F.4.2 SETTING DETAILS

  • Herd effects are measured by post score Si = Ui − Di and GPT-4o-rated disagree scores, comparing means and 95% confidence intervals across up-treated, down-treated, and control groups.The human-study replication uses popularity-based recommendations, controlled posting, restricted follow and mute actions, cached top-300 posts, 0.1 activation probability, and 40 time steps.Source: §F.4.1 · p. 34, §F.4.2 · pp. 34–35

F.4.2 SETTING DETAILS

  • The experiment used 100, 1k, or 10k agents, with 0.1 activation probability per step, five randomly sampled recommended posts per activated agent, and 30 total time steps.Recommendation caches contained 50, 500, or 5,000 posts, while controlled users created 30, 300, or 3k posts per step until datasets of 219, 2,191, or 21,919 posts were completed.Source: §F.4.2 · p. 35

F.4.3 EXAMPLES OF RESULTS

OASIS supports large-scale discussion of counterfactual posts, where agent viewpoints shift progressively toward rejection and group-level factual correction. The section also provides paired truth and misinformation examples across technology, health, education, and entertainment.

F.4.3 EXAMPLES OF RESULTS

  • 10,000 agents gradually shifted toward opposing counterfactual content, achieving factual correction at the group level.Agents interacted by posting comments and liking or disliking others’ comments.Source: §F.4.3 · p. 35
  • Agent viewpoints progressed from surprise and curiosity to partial opposition and finally complete rejection as interactions increased.Source: Figure 16 · p. 35

G.0.1 TRUTH AND MISINFORMATION PAIRS

  • Truth–misinformation pairs cover technology, health, education, and entertainment examples, including claims about drones, exercise, scholarships, and Marvel films.Source: §G.0.1 · pp. 35–36

H LIMITATIONS & FUTURE DIRECTIONS

OASIS’s future development is bounded by simplified recommendation, abstracted users, incomplete platform actions, multimodal and runtime constraints, and ethical risks. The authors identify opportunities for using the simulator to evaluate recommendation systems and train LLMs with network feedback.

H LIMITATIONS & FUTURE DIRECTIONS

I SOCIAL IMPACT AND ETHICAL CONSIDERATIONS

  • Using LLM agents and real-world data raises concerns about result fidelity, bias, privacy, and misuse for manipulating online discourse or misinformation campaigns.Source: I SOCIAL IMPACT AND ETHICAL… · p. 37

Related papers

Loading 2411.11581v5…