Decoding Echo Chambers: LLM-Powered Simulations Revealing Polarization in Social Networks

TL;DR

This paper tackles echo chambers and polarization in social networks by introducing a language-based Social Simulation Framework (SSF) that employs LLM agents operating on graph-structured networks. Agents reason with memory and natural language updates, generating opinions $v_i \in [-2,2]$ over $T$ days and interacting via graph-aware recommendations; the approach is contrasted with numeric BCM and Friedkin-Johnsen dynamics to highlight the benefits of text-based reasoning. It demonstrates that SSF captures polarization and echo-chamber phenomena in small-world and scale-free networks, while random graphs do not, and it validates two language-driven mitigation strategies (Active and Passive Nudges) that reduce extremity and inter-neighbor alignment. The work emphasizes explainability through conversational updates and offers practical insights for mitigating polarization, while acknowledging limitations in scale ($N=50$) and potential LLM biases, suggesting pathways to larger, more diverse simulations. Overall, the approach provides a new, interpretable avenue for studying social polarization and evaluating interventions in language-rich social environments, with implications for governance and platform design.

Abstract

The impact of social media on critical issues such as echo chambers needs to be addressed, as these phenomena can have disruptive consequences for our society. Traditional research often oversimplifies emotional tendencies and opinion evolution into numbers and formulas, neglecting that news and communication are conveyed through text, which limits these approaches. Hence, in this work, we propose an LLM-based simulation for the social opinion network to evaluate and counter polarization phenomena. We first construct three typical network structures to simulate different characteristics of social interactions. Then, agents interact based on recommendation algorithms and update their strategies through reasoning and analysis. By comparing these interactions with the classic Bounded Confidence Model (BCM), the Friedkin Johnsen (FJ) model, and using echo chamber-related indices, we demonstrate the effectiveness of our framework in simulating opinion dynamics and reproducing phenomena such as opinion polarization and echo chambers. We propose two mitigation methods, active and passive nudges, that can help reduce echo chambers, specifically within language-based simulations. We hope our work will offer valuable insights and guidance for social polarization mitigation.

Figures (5)

• Figure 1: Our language-based simulation provides an explainable and dynamic environment compared with numeric simulation for echo chamber study.
• Figure 2: Our framework is evaluated on three different network structures that mimic real-world observations. Each agent is initialized with personal information, dual memory, and a reasoning process. Through random or recommendation-based interactions, they update their opinions each day.
• Figure 3: We present projection and curve graphs from BCM and our SSF simulations under three settings to demonstrate our framework's macro-level effectiveness. The main conclusions are similar: small-world and scale-free networks lead to severe echo chamber effects, while the random graph does not.
• Figure 4: Case study: The same person in a non-mitigation environment and during our mitigation operation. His opinion is more peaceful and less aggressive in our setting.
• Figure 5: We propose two nudge operations to mitigate echo chambers and polarization. Compared with the default setting in Figure \ref{['fig:main']}, it is evident that both phenomena are better alleviated.