CoevolveSim: A Simulation Framework for Studying Belief Coevolution in Large Language Model Social Networks

CoevolveSim is an agent-based simulation framework designed to study how beliefs coevolve in social networks of large language models (LLMs). This project explores how social interactions alter the belief dynamics of LLMs and how expert agents influence collective outcomes.

• Original Authors/Maintainers: Germans Savcisens, Samantha Dies, Courtney Maynard, Tina Eliassi-Rad (Northeastern University, USA)
• Source Platform: GitHub
• Original Project Name: coevolve-sim
• Original Link: https://github.com/carlomarxdk/coevolve-sim
• Release Year: 2025 (corresponding DOI: 10.5281/zenodo.17875304)

---

With the widespread application of large language models (LLMs) across various domains, people are increasingly aware that these models do not exist in isolation—they are deployed in diverse social environments such as social media, forums, and customer service systems, where they interact and influence each other. However, the scientific understanding of how beliefs form, spread, and evolve among LLMs in social networks remains limited.

CoevolveSim was created to address this research gap. It provides a systematic simulation framework that allows researchers to observe and analyze the process of belief coevolution among LLM agents in social networks. This tool is of great significance for understanding the collective behavior of AI systems, information dissemination mechanisms, and potential bias amplification effects.

---

Belief coevolution refers to the phenomenon where the belief systems of multiple agents influence each other and change together during continuous interactions. This stands in sharp contrast to traditional isolated inference—within a social network, each agent's output becomes input for others, forming complex feedback loops.

CoevolveSim specifically focuses on two types of LLM agents:

• Generalist Models: Possess broad knowledge but lack depth in specific domains
• Specialist Models: Have expertise in specific domains

By comparing the performance of these two types of agents in social networks, researchers can explore the mechanisms of professional knowledge dissemination and the impact of expert opinions on collective decision-making.

---

Each simulation run in CoevolveSim follows a concise yet powerful cycle:

Agents first form initial beliefs about a certain statement. This step simulates the agent's initial judgment on a specific proposition based on its own knowledge base.

Agents receive belief summaries from neighboring nodes in the social network. This simulates the process of individuals acquiring others' opinions in a social environment. Network structures (such as random networks, small-world networks, etc.) significantly influence the flow path of information.