YuLan-OneSim: Towards the Next Generation of Social Simulator with Large Language Models

TL;DR

YuLan-OneSim presents a scalable, LLM-driven social simulator with code-free scenario construction, a large default repository, evolvable simulations via a VR2T feedback loop, and a distributed Master-Worker architecture supporting up to 100,000 agents. Central to the framework are four subsystems—scenario auto-construction, simulation, feedback-driven evolution, and an AI social researcher—that together enable an end-to-end social science research loop. The paper demonstrates high-quality automatically generated scenarios (BG-Rating 4.82, C-Rating 4.20), reliable simulation behavior through social-theory verification and real-world data fitting, and strong but improvable performance of the AI social researcher. The work provides a practical path toward automated, large-scale, cognitively grounded social simulations with potential for broad impact across social science research and policy analysis. The authors also release the codebase, inviting the community to extend scenarios, theory coverage, and multi-modal inputs for richer simulations.

Abstract

Leveraging large language model (LLM) based agents to simulate human social behaviors has recently gained significant attention. In this paper, we introduce a novel social simulator called YuLan-OneSim. Compared to previous works, YuLan-OneSim distinguishes itself in five key aspects: (1) Code-free scenario construction: Users can simply describe and refine their simulation scenarios through natural language interactions with our simulator. All simulation code is automatically generated, significantly reducing the need for programming expertise. (2) Comprehensive default scenarios: We implement 50 default simulation scenarios spanning 8 domains, including economics, sociology, politics, psychology, organization, demographics, law, and communication, broadening access for a diverse range of social researchers. (3) Evolvable simulation: Our simulator is capable of receiving external feedback and automatically fine-tuning the backbone LLMs, significantly enhancing the simulation quality. (4) Large-scale simulation: By developing a fully responsive agent framework and a distributed simulation architecture, our simulator can handle up to 100,000 agents, ensuring more stable and reliable simulation results. (5) AI social researcher: Leveraging the above features, we develop an AI social researcher. Users only need to propose a research topic, and the AI researcher will automatically analyze the input, construct simulation environments, summarize results, generate technical reports, review and refine the reports--completing the social science research loop. To demonstrate the advantages of YuLan-OneSim, we conduct experiments to evaluate the quality of the automatically generated scenarios, the reliability, efficiency, and scalability of the simulation process, as well as the performance of the AI social researcher.

Figures (74)

• Figure 1: The overall framework of our simulator. The upper subfigure highlights the key features of our simulator (i.e., logical characteristics), while the lower subfigure illustrates its core subsystems (i.e., physical implementations). The correspondences between the features and their supporting subsystems are labeled by solid arrows.
• Figure 2: A toy example of the progressive scenario construction in the job market simulation: from the logical level to the data structure level, and finally to the concrete data level.
• Figure 3: Illustration of the working process of the AI Social Researcher. There are two main modules: the Experiment design module and the Report generation module. YuLan-OneSim serves as the central tool for executing simulations and enabling a complete loop of social science research.
• Figure 4: Illustration on the distribution of various error types—including logical errors, syntax errors, and robustness issues—across different domains.
• Figure 5: Cultural Similarity Maps across simulation stages. The four panels show the evolution of cultural similarity between adjacent agents at rounds 0, 25, 50, and 100. Darker connections indicate lower cultural similarity, revealing the formation of distinct cultural regions over time. The color gradient corresponds to similarity percentages from 0% (darkest) to 100% (lightest).