MTOS: A LLM-Driven Multi-topic Opinion Simulation Framework for Exploring Echo Chamber Dynamics

Dingyi Zuo, Hongjie Zhang, Jie Ou, Chaosheng Feng, Shuwan LiuarXiv 2025 · DOI · arXiv

TL;DR

MTOS is a new simulation framework that uses Large Language Models (LLMs) to model how opinions evolve across multiple interrelated topics simultaneously. By accounting for "attention dispersion" and cognitive fatigue, the researchers demonstrate that introducing unrelated or negatively correlated topics can actually break the feedback loops that create social media echo chambers.

What Problem Does This Paper Address?

Most research into "echo chambers"—those digital bubbles where we only hear opinions that match our own—treats opinion as a single numerical value on a single topic (e.g., "Are you for or against Topic X?").

However, real-world social dynamics are far messier. Our stance on climate change might be linked to our views on economic policy, and our "cognitive budget" is limited; if we are arguing about sports, we have less energy to polarize over politics. Traditional numerical models lack the linguistic nuance to capture these cross-topic interactions, while earlier LLM-based simulations usually restricted agents to discussing just one thing at a time. MTOS aims to bridge this gap by simulating a multi-topic environment where topics compete for a user's limited attention.

Methods

The researchers built the Multi-topic Opinion Simulation (MTOS) framework using the Python library Mesa.

1. The Agents

The simulation features 50 agents, each powered by the Qwen-2.5 (7B) LLM. To make them "human-like," each agent is assigned:

  • Demographics: Age, gender, and education.
  • Personality: Randomly assigned traits based on the "Big Five" model (Openness, Conscientiousness, etc.).
  • Dual-Layer Memory: A short-term buffer for recent interactions and a long-term memory for historical context.

2. Interaction Mechanisms

The framework tests two ways agents decide who to talk to:

  • Multidimensional Hegselmann–Krause (HK): A numerical approach where agents talk to neighbors whose average belief scores across all topics are within a certain "tolerance" threshold.
  • Semantic Matching: A more advanced LLM-driven approach where the agent reads a "structured prompt" about a neighbor's traits and beliefs and decides—using reasoning—whether to engage.

3. Dynamic Topic Selection

Instead of talking about everything at once, agents use a recommendation system that balances:

  • Topic Heat: How popular a topic is in the global network.
  • Topic Fatigue: An exponential decay function that makes an agent less likely to discuss a topic they’ve already spent too much time on.

Datasets

The paper does not rely on a static external dataset for training. Instead, it uses the Qwen-2.5 (7B) pre-trained model as the cognitive engine. The "data" is generated synthetically through the simulation process. The researchers initialized the environment with 50 agents and ran simulations over 30 "days" (rounds) to observe the emergence of patterns.

Key Insights and Ideas

  • The "Attention Dispersion" Effect: One of the most striking findings is that simply adding irrelevant topics to a social network can mitigate echo chambers. Because agents have limited cognitive resources, the presence of "noise" or unrelated discussions prevents them from hyper-focusing and polarizing on a single controversial issue.
  • Correlation is Key:
    • Positively correlated topics (e.g., two topics that align with the same worldview) act as force multipliers for echo chambers.
  • Negatively correlated topics (where supporting one implies opposing the other) create "cognitive friction" that inhibits polarization.
  • Reasoning vs. Rules: Unlike traditional models that use hard-coded math to update opinions, MTOS agents use the LLM to "reflect" on conversations. They can decide to change their mind based on the quality of an argument and their own personality, leading to more stable and realistic system behavior.

Key Results

The researchers compared MTOS against a single-topic baseline (SSF).

  • Echo Chamber Index (ECI): In single-topic environments, the ECI rose steadily, indicating rapid polarization. In the MTOS multi-topic environment, the ECI remained significantly lower, suggesting that multi-topic interaction is a natural defense against radicalization.
  • Neighbor Correlation Index (NCI): MTOS showed a much lower NCI than single-topic models. This means agents in MTOS maintained more diverse viewpoints even when surrounded by neighbors, effectively resisting the "homophily" trap where everyone eventually thinks the same.
  • Ablation Success: Removing the "belief decay" or "topic selection" mechanisms caused the simulation to revert to more polarized, less realistic states, proving that cognitive fatigue and topic competition are essential for modeling human-like social networks.

Limitations and Future Work

While 50 agents is a step up from many previous LLM studies, it is still a tiny fraction of a real social network. The authors acknowledge that future work needs to:

  • Scale Up: Test the framework with hundreds or thousands of agents.
  • Expand Topic Range: Introduce a wider variety of value orientations and more complex inter-topic dependencies.
  • LLM Optimization: Fine-tune the underlying models to ensure they don't have inherent biases that might skew the simulation results toward neutrality or specific viewpoints.

Summary updated

Abstract

The polarization of opinions, information segregation, and cognitive biases on social media have attracted significant academic attention. In real-world networks, information often spans multiple interrelated topics, posing challenges for opinion evolution and highlighting the need for frameworks that simulate interactions among topics. Existing studies based on large language models (LLMs) focus largely on single topics, limiting the capture of cognitive transfer in multi-topic, cross-domain contexts. Traditional numerical models, meanwhile, simplify complex linguistic attitudes into discrete values, lacking interpretability, behavioral consistency, and the ability to integrate multiple topics. To address these issues, we propose Multi-topic Opinion Simulation (MTOS), a social simulation framework integrating multi-topic contexts with LLMs. MTOS leverages LLMs alongside short-term and long-term memory, incorporates multiple user-selection interaction mechanisms and dynamic topic-selection strategies, and employs a belief decay mechanism to enable perspective updates across topics. We conduct extensive experiments on MTOS, varying topic numbers, correlation types, and performing ablation studies to assess features such as group polarization and local consistency. Results show that multi-topic settings significantly alter polarization trends: positively correlated topics amplify echo chambers, negatively correlated topics inhibit them, and irrelevant topics also mitigate echo chamber effects through resource competition. Compared with numerical models, LLM-based agents realistically simulate dynamic opinion changes, reproduce linguistic features of news texts, and capture complex human reasoning, improving simulation interpretability and system stability.
BibTeX
@misc{zuo2025mtos,
  author = {Dingyi Zuo and Hongjie Zhang and Jie Ou and Chaosheng Feng and Shuwan Liu},
  title = {{MTOS: A LLM-Driven Multi-topic Opinion Simulation Framework for Exploring Echo Chamber Dynamics}},
  year = {2025},
  eprint = {2510.12423},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.12423},
  howpublished = {arXiv},
}
Figure