SocioVerse - An open-source social simulation world model developed by Fudan University in collaboration with Xiaohongshu and other institutions.
SocioVerse is a social simulation world model jointly launched by Fudan University, Shanghai Institute of Innovation and Entrepreneurship, University of Rochester, and Xiaohongshu. It is based on a Large Language Model (LLM) driven intelligent agent and includes 10 million real users...
What is SocioVerse?
SocioVerse is a social simulation world model jointly launched by Fudan University, Shanghai Institute of Innovation and Entrepreneurship, University of Rochester, and Xiaohongshu. Based on Large Language Model (LLM)-driven agents and a data pool containing 10 million real users, it constructs a large-scale social simulation framework "aligned" with the real world. The model is based on four modules: social environment, user engine, scene engine, and behavior engine, enabling it to model the behavioral patterns of target groups, predict the evolution trends of group events, and assist in major decision-making. SocioVerse demonstrates high-precision alignment in multiple scenarios, including news hotspot dissemination and socio-economic surveys, providing a powerful tool for interdisciplinary research in computational social sciences.
SocioVerse's main functions
- High-precision social simulationTo construct simulated environments aligned with the real world, simulate large-scale group behavior, and predict the evolutionary trends of social events.
- Multi-scenario applicationsIt supports various social research scenarios, such as news dissemination, socio-economic surveys, and political election prediction, helping researchers and policymakers gain insights into group behavior.
- User profiling and behavior generation: Construct complex target user profiles and generate behavioral patterns that conform to user characteristics based on intelligent agents.
- Dynamic environment updatesBased on real-time updates of social events, social statistics, and personalized content, the simulated environment keeps pace with the real world.
SocioVerse's technical principles
- Social Environment ModuleThe system provides the simulation with up-to-date social context information, including real-time events, social statistics, and personalized content. Event updates build a timestamped news event database for the agent to retrieve and reference. Social statistics provide structured data such as population distribution and urban structure, making agent behavior more consistent with group characteristics. Preference content is pushed to the agent based on a recommendation system, enhancing the diversity and personalization of behavior generation.
- User EngineThe system constructs target user profiles based on real user data to ensure that the characteristics of the simulated intelligent agents are consistent with the real-world distribution. The user pool contains data from 10 million users across multiple social media platforms. User tags combine hard labels (such as gender and age) and soft representation vectors, generating high-precision user profiles based on large language model annotation and manual verification.
- Scene EngineAligning simulated scenarios with real-world scenarios, designing corresponding interaction structures, and extending to large-scale groups based on population distribution. Questionnaire surveys, based on a one-to-many single-round structure, collect opinions from a large sample. In-depth interviews utilize one-to-one multi-round interactions to uncover respondents' attitudes and motivations. Behavioral experiments, based on one-to-many or many-to-many structures, observe individual and group decision-making behaviors. Social media interactions, based on many-to-many dynamic posting and commenting, simulate information diffusion and network influence.
- Behavior EngineCombining user profiles, scene structure, and social context, intelligent agents are driven to generate reasonable simulated behaviors. Large-scale model intelligent agents include general-purpose LLMs (such as GPT and Qwen), expert LLMs (fine-tuned for specific domains), and domain LLMs (handling complex tasks). Traditional modeling intelligence is based on rules or mathematical models, suitable for modeling low-influence edge users, and has the advantage of computational efficiency.
SocioVerse's project address
- Project official website:http://www.fudan-disc.com/socioverse/
- GitHub repository:https://github.com/FudanDISC/SocioVerse
- HuggingFace model library:https://huggingface.co/datasets/Lishi0905/SocioVerse
- arXiv technical paper:https://arxiv.org/pdf/2504.10157
Application scenarios of SocioVerse
- Political election predictionsSimulate the election process, predict candidate support rates and election results, and provide data support for political campaign strategies.
- Breaking News Feedback AnalysisSimulate public reactions to major news events to help assess public attitudes and develop response strategies in advance.
- National Economic SurveySimulates residents' consumption behavior and economic decision-making to provide a reference for economic policy formulation and market research.
- Social Policy AssessmentPredicting the effects of social policies and public reactions can help optimize policy design.
- Information dissemination and public opinion analysisSimulates the spread of information and the evolution of public opinion in social networks, providing analytical tools for public opinion management and information governance.