SPARK: Simulating the Co-evolution of Stance and Topic Dynamics in Online Discourse with LLM-based Agents

Bowen Zhang, Yi Yang, Fuqiang Niu, Xianghua Fu, Genan Dai, Hu HuangEMNLP 2025

TL;DR

This paper introduces SPARK, a novel LLM-based multi-agent simulation framework that models how topics and user stances evolve together in online discussions. By simulating agents with distinct personalities and memory, SPARK offers a more interpretable and fine-grained way to understand and potentially mitigate online polarization.

What Problem Does This Paper Address?

Online discussions are dynamic, with topics constantly shifting and users adopting various stances. Existing models often treat topic evolution and stance dynamics separately, leading to abstractions that lack real-world fidelity and interpretability. This disconnect makes it difficult to understand the complex interplay between how discussions change and how people's opinions evolve in response. The authors aim to bridge this gap by creating a simulation that captures these co-evolutionary mechanisms at an agent level, offering insights into phenomena like polarization and misinformation spread.

Methods

The core of SPARK is a multi-agent simulation where each agent is an LLM persona. These agents are equipped with:

  • Persona Initialization: Each agent has a unique profile, including demographic and psychological traits (based on the Big Five model), which influences their initial stance and susceptibility to influence.
  • Dual Memory System: Agents maintain both short-term memory (for daily interactions) and long-term memory (for historical context), allowing for nuanced, context-aware stance updates.
  • Text-based Stance Update: Instead of simple numerical values, agents express and update their stances through natural language (tweets), making the simulation more interpretable.
  • Topic Evolution Simulator (TES): This component manages the emergence and propagation of new subtopics based on agent interactions and hashtags, creating a dynamic topic tree.

The simulation proceeds in daily rounds where agents converse, update their memories and stances, generate tweets, and introduce new topics, mimicking real-world online discourse.

Datasets

The paper does not explicitly mention training datasets. For evaluation and simulation, they used five diverse topic domains: politics, technology, society, medicine, and science. The agent personas are initialized using a name dataset and the Big Five personality traits.

Key Insights and Ideas

  • Co-evolution is Key: The most significant insight is the strong, bidirectional reinforcement between topic evolution and stance change. Shifts in stance actively promote new topic generation, and vice-versa.
  • Personality Matters: Agent personality traits (e.g., low neuroticism) influence stance persistence, suggesting that personalized interventions might be more effective than uniform strategies for managing online discourse.
  • Domain-Specific Dynamics: Different topics exhibit distinct evolution patterns. Politics and science tend towards strong consensus or polarization, while technology remains more diverse and contentious.
  • Memory is Crucial: The ablation study highlights that both short-term and long-term memory components are essential for maintaining realistic and diverse stance dynamics. Without them, simulations become monotonous or fail to sustain meaningful evolution.

Key Results

  • Reproduces Real-World Patterns: SPARK successfully replicates empirical patterns like rapid topic innovation in technology and domain-specific stance polarization.
  • Quantitative Correlation: A strong positive correlation (r=0.88) was found between stance change and the emergence of new topics.
  • Personality Effects: Credulous agents (higher Agreeableness and Neuroticism) drove higher topic novelty (TD) compared to Skeptical agents.
  • Stance Diversity: The full architecture achieved the highest stance diversity (0.504), with memory components being critical for this.
  • Domain Differences: Politics and Science showed high Stance Bias (OB) and low Stance Diversity (OD), indicating strong consensus/polarization. Technology and Society had more moderate bias and higher diversity.

Limitations and Future Work

  • Simplified Environment: The current simulation environment is relatively simplified and doesn't fully capture the complexity of real-world platforms like Facebook or Twitter.
  • Limited Factors: The study primarily focuses on Big Five personality traits and topic types, not incorporating other potential influencing factors.
  • Future Directions: The authors plan to incorporate richer network structures and external information flows to enhance the simulation's realism and advance the study of complex social discourse. They also aim to conduct simulations in environments that more closely resemble actual social platforms to improve practical significance.

Summary updated

BibTeX
@inproceedings{zhang2025spark,
  author = {Bowen Zhang and Yi Yang and Fuqiang Niu and Xianghua Fu and Genan Dai and Hu Huang},
  title = {{SPARK: Simulating the Co-evolution of Stance and Topic Dynamics in Online Discourse with LLM-based Agents}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure