Simulation Achieves Realistic Online Social Networks, Enabling Disinformation Campaign Analysis

Understanding the spread of misinformation online presents a significant challenge for researchers, often hampered by data limitations and ethical concerns surrounding real-world social networks. Alejandro Buitrago López, Alberto Ortega Pastor, and David Montoro Aguilera, from the University of Murcia, along with their colleagues, address this problem by developing a novel simulation framework that generates realistic and interpretable online social environments. Their work overcomes limitations in existing models by creating synthetic networks populated by agents possessing distinct personality traits and predictable behaviours, allowing for detailed study of information flow. Crucially, the team incorporates a system for simulating coordinated disinformation campaigns, and validates the framework’s realism through rigorous analysis of network structure, agent behaviour, and the language used in generated posts, offering a powerful new tool for investigating the dynamics of online information and the impact of malicious actors.

LLMs, Agents, and Social Simulations

This work explores the intersection of Large Language Models (LLMs), agent-based modeling, and social network analysis, offering a comprehensive overview of current research and potential applications. A central theme is the development of autonomous agents powered by LLMs like Gemini, designed to simulate human-like interactions in digital environments, utilizing technologies such as Sentence-BERT for embeddings and Ollama for LLM execution. The research leverages computational social science and agent-based modeling to simulate social phenomena, focusing on network structure, centrality measures, and user behavior on platforms like Twitter and Mastodon. Key technologies include LLMs, agent-based modeling, sentence embeddings, and vector databases, alongside concepts like Markov Chains and social graphs. This research is valuable for researchers in LLMs, computational social science, and social network analysis, providing a starting point for new projects and a snapshot of current trends. The concepts and technologies described can be used to build simulations of social phenomena, analyze social media data, and develop LLM-powered applications.

Realistic Agent-Based Social Network Simulation

The research team engineered a novel simulation framework to model online social networks, prioritizing realism and interpretability, and addressing limitations in existing systems. This work pioneers a method for creating synthetic social networks populated by agents with demographic-based personality traits and finite-state behavioral automata, enabling the simulation of nuanced and understandable actions. A generative module, powered by a large language model, produces context-aware social media posts, ensuring each agent’s contributions align with its established profile and remembered experiences, enhancing behavioral authenticity. In parallel, the study implemented a “red” module inspired by DISARM workflows, enabling the orchestration of disinformation campaigns executed by malicious agents.

The system delivers a Mastodon-based visualization layer, providing real-time inspection of agent activity and facilitating post-hoc validation of simulated interactions. Rigorous evaluation using topological metrics and language model-based content assessments demonstrates structural, behavioral, and linguistic realism, reproducing key characteristics of real-world social networks, including small-world structures and patterns of homophily. The framework successfully simulates up to one million agents, offering a customizable and controllable environment for in-depth analysis of social network phenomena and the effects of coordinated disinformation efforts.

Realistic Synthetic Social Networks for Disinformation Study

Scientists have developed a novel simulation framework capable of generating realistic and interpretable synthetic social networks, addressing limitations in existing research tools. The work centers on creating virtual online social networks populated by agents possessing demographic-based personality traits and defined behavioral patterns, allowing for detailed study of information flow and the impact of disinformation. A key component is a generative module, powered by a large language model, which produces context-aware social media posts aligned with each agent’s individual profile and evolving memory, ensuring realistic content creation. Experiments demonstrate the framework’s ability to create networks exhibiting structural, behavioral, and linguistic realism, validated through topological metrics and language model assessments.

The team measured network formation, revealing the system’s capacity to reproduce patterns like homophily and triadic closure, resulting in small-world structures mirroring real-world social connections. The framework incorporates “red” agents, modeled on DISARM-inspired workflows, to orchestrate disinformation campaigns, enabling systematic evaluation of countermeasures. The simulation generated over 10,000 agents, simulating polarization, message diffusion, policy experiments, and disasters within a controlled environment, and produced over 7.7 million posts, with agents exhibiting behaviors such as increased emoji usage and a tendency to engage with abusive content. Measurements confirm that the generated networks display broader connectivity and lower clustering compared to existing platforms, with abusive agents occupying central positions, providing valuable insights into the dynamics of online harm. A Mastodon-based visualization layer allows for real-time inspection and post-hoc validation of agent activity, enhancing the framework’s utility as a reproducible testbed for analyzing emergent social phenomena and evaluating intervention strategies.

Realistic Social Network Simulation with Agent Profiles

This research presents a new simulation framework for modelling online social networks, designed to overcome limitations in existing approaches, such as a lack of realism and difficulty in interpreting results. The framework creates synthetic networks populated by agents possessing personality traits and predictable behaviours, allowing researchers to study information flow and the impact of disinformation campaigns in a controlled environment. A key innovation lies in combining agent-based modelling with large language models, which generate realistic social media posts aligned with each agent’s established profile and memory, while maintaining interpretable actions governed by defined rules. The resulting simulations demonstrate structural, behavioural, and linguistic realism, reproducing key characteristics of real-world social platforms. The framework’s modular design allows for integration with existing data analytics tools, broadening its potential applications for research into areas like bot detection and natural language processing. While acknowledging the need for further improvements in scalability and content evaluation, the team highlights the value of synthetic simulations as a safe and transparent means of investigating online discourse and the spread of misinformation.