Using social simulation with artificial intelligence agents instead of traditional polls

What if large language models (LLM) can simulate public opinion, reactions and social dynamics, without the need for traditional polls?

University researchers Food China a framework called Socioverse They have developed that a social simulation model is based on artificial intelligence agents. This framework can replace traditional methods such as polls, interviews, and observations, often accompanied by challenges such as high costs, sample size limitations, and ethical concerns.

Socioverse: Framework for Social Simulation

Socioverse Includes four main components and a database of 1 million real users Is.

• The component of the social environment: Inserts up -to -date external information to simulation.
• Motor users (User Engine): Provides real areas of users.
• Scenario motor (Scenario Engine): Adapts the simulation to the real world.
• The motor of behaviors (Behavior Engine)A: It reproduces human behaviors by agents.

This framework allows researchers to observe the patterns of micro-Level and predict future social dynamics.

Experiments and results

To evaluate this framework, researchers conducted simulation tests in various fields such as politics, news and economics. Models like LLAMA-3-70B-Instruct, Qwen2.5-72B-Instruct, Deepseek-V3, Gpt-4o Mini, Gpt-4O And Deepseek-R1-671B They were used for these experiments.

1. Predicting US Presidential Election Results:
   LLM agents were able to over 1% Precision Predict the results of voting in states. Models Gpt-4o-mini And QWen2.5-72B Had competitive performance, while the model Deepseek-R1-671B Due to “over -thinking”, the results had lower results in terms of precision.
2. Public reaction to important news:
   Simulation of public reaction to publishing Chatgpt Done. Models Gpt-4O And QWen2.5-72B They showed the most coordination with the real views of society.
3. Investigating National Economic Behaviors in China:
   In a test of economic behavior simulation through a national poll in China, the model Llama-3-70B It was better to reproduce people’s habits better than other models.

Capabilities of AI models in social simulation

“Our findings show that advanced Language models (LLM) have a significant ability to simulate human reactions in complex social areas, although there are still gaps between simulated responses and the actual observed results,” the research article said.

The future of social simulations

Researchers intend to expand the scope of the scenarios under study to improve the simulation capabilities of large language models (LLM). This innovation can create a revolution in the methods of studying social behaviors and be an effective and efficient replacement for traditional social surveys and research.