Unveiling the Truth and Facilitating Change: Towards Agent-based Large-scale Social Movement Simulation

TL;DR

This work addresses the challenge of simulating large-scale social movement dynamics on social media by introducing HiSim, a hybrid framework that models a small set of LLM-driven core users and a large population of ABM-driven ordinary users within a Twitter-like environment. It formalizes opinion dynamics, integrates LLM-powered agents with memory and planning, and defines a two-tier simulation process to capture micro- and macro-level phenomena. A dedicated benchmark, SoMoSiMu-Bench, enables data-driven evaluation across three real-world movements, with extensive experiments demonstrating improved micro/macro alignment and scalable performance, along with insights on echo chambers and interventions. The approach offers a practical, scalable path for forecasting public opinion and evaluating interventions to promote constructive online discourse, while acknowledging data privacy and potential biases in LLM-generated behavior.

Abstract

Social media has emerged as a cornerstone of social movements, wielding significant influence in driving societal change. Simulating the response of the public and forecasting the potential impact has become increasingly important. However, existing methods for simulating such phenomena encounter challenges concerning their efficacy and efficiency in capturing the behaviors of social movement participants. In this paper, we introduce a hybrid framework HiSim for social media user simulation, wherein users are categorized into two types. Core users are driven by Large Language Models, while numerous ordinary users are modeled by deductive agent-based models. We further construct a Twitter-like environment to replicate their response dynamics following trigger events. Subsequently, we develop a multi-faceted benchmark SoMoSiMu-Bench for evaluation and conduct comprehensive experiments across real-world datasets. Experimental results demonstrate the effectiveness and flexibility of our method.

Figures (6)

• Figure 1: An illustration of user interactions and attitude changes after a trigger event happens. Users can take actions such as posting and retweeting according to their traits, and their generated content will be stored in the Twitter timeline and fed to their connected users. Users can change attitudes once perceive others' opinions.
• Figure 2: The proposed framework architecture. The bottom part illustrates the architecture of core users and the mechanism for ordinary users. The top part presents the simulation process. At each round, core user agents take action by generating textual responses based on contextual information, and their attitudes are conveyed to ordinary users after postprocessing, while ordinary users communicate using attitude scores directly.
• Figure 3: (a) System metrics when simulating with varying numbers of agents (better viewed in color); (b) Running Efficiency with varying numbers of agents.
• Figure 4: (a) Similarity of content of consumption and production on Metoo simulation; (b) Similarity of content of consumption and production on Roe simulation. BLM is not reported since it is a partial phrase.
• Figure 5: The data processing process.