What is YuLan-OneSim
YuLan-OneSim is a new type of social simulator launched by the RUC-GSAI team at the Gaoling School of Artificial Intelligence, Renmin University of China. It simulates human social behavior based on large language model (LLM) Agents, builds simulation scenarios without programming, and generates code based on natural language interaction. YuLan-OneSim provides more than 50 default scenarios covering 8 major social science fields, supports large-scale simulations with up to 100,000 Agents, and automatically optimizes LLMs based on external feedback. YuLan-OneSim has AI social researcher capabilities, automatically generating reports from research topics and completing the entire social science research cycle. YuLan-OneSim promotes deep integration of social science and AI, providing a powerful tool for social science research.
Main Features of YuLan-OneSim
- Build simulation scenarios without programming: Users describe and refine simulation scenarios through natural language interaction, and the system automatically generates the corresponding execution code, reducing the requirement for programming skills.
- Rich default scenario library: Provides more than 50 default simulation scenarios covering 8 major social science fields, including economics, sociology, political science, etc., providing researchers with rich research materials.
- Evolvable simulation: Automatically optimizes LLMs based on external feedback, improving the accuracy and reliability of simulations.
- Large-scale simulation capability: Based on a distributed architecture, supports simulations with up to 100,000 Agents.
- AI social researcher: Automatically transforms research topics into specific simulation scenarios and generates research reports, automating the research process from proposing research topics to generating reports.
Technical Principles of YuLan-OneSim
- Scenario formalization: Based on the Overview, Design Concepts, and Details (ODD) protocol, transforms user requirements into structured scenario descriptions.
- Behavior graph construction: Based on the ODD protocol, extracts Agent types and behavior logic, generates behavior graphs, and defines the interaction logic between Agents.
- Code generation: Generates executable simulation code according to the behavior graph, using a modular code generation method to ensure the accuracy and maintainability of the code.
- Scenario specification: Generates environment data, Agent configuration data, and Agent relationship data to ensure the completeness and consistency of the simulation.
- Simulation subsystem: Agents are composed of multiple modules, including configuration, memory, planning, and action modules, supporting a high degree of customization. It uses an event-driven asynchronous event bus to support parallel computing and efficient event processing. Based on a master-slave node architecture, it supports parallel computing for large-scale Agents and optimizes communication efficiency and resource allocation.
- Feedback-driven evolution subsystem: Based on the multi-agent framework (Verifier–Reasoner–Refiner–Tuner, VR²T), evaluates and optimizes simulation results. The prompt-response pairs generated by the system are verified, reasoned, corrected, and fine-tuned to improve LLM performance.
- AI social researcher subsystem: The experimental design module transforms research topics into specific simulation scenarios, including generating candidate research questions, evaluating scenario feasibility, and generating the ODD protocol. The report generation module generates detailed analysis reports based on simulation results, including data interpretation, report structure generation, report writing, and report review.
Project Address of YuLan-OneSim
- GitHub Repository：https://github.com/RUC-GSAI/YuLan-OneSim
- arXiv Technical Paper：https://arxiv.org/pdf/2505.07581
Application Scenarios of YuLan-OneSim
- Economics: Simulate market dynamics such as labor matching, auction markets, bank reserves, etc., and analyze the impact of policies and mechanisms.
- Sociology: Study social phenomena such as cultural capital, information diffusion, social norms, etc., to understand the interaction between social structure and individual behavior.
- Political Science: Simulate voter behavior, policy implementation, electoral polarization, etc., and analyze the social impact of institutions and policies.
- Psychology: Simulate cognitive dissonance, emotional contagion, conformity behavior, etc., to explore the internal mechanisms of psychological behavior.
- Public Health: Simulate infectious disease transmission, health inequality, community health mobilization, etc., and evaluate the effectiveness of intervention measures.