Opinion Polarization in LLM-Based Social Networks: Manipulation and Mitigation

Society, culture, and collective behavior2026arXivApproved editorial review

Authors: Ali Safarpoor Dehkordi, Mohammad Shirzadi, Ahad N. Zehmakan

Keywords: Multi-agent systems, Collective behavior, Social simulation

Source: Open primary source (opens in a new tab)

3
Authors
4
Findings
5
Limitations
3
Evidence

Editorial summary

English

Agents with numeric opinion, stubbornness, and activity generate posts and update positions on synthetic HRG and Twitter/Reddit graphs. Random, degree, centrality, and community selection are compared; adversaries are persistent or susceptible; reactive moderators and exposure, feed, filtering, activity, and connection interventions are tested. Results average 10 runs.

Synthetic networks and graphs derived from Twitter and Reddit, with GPT-4.1-mini as the main model and GPT-4o-mini and DeepSeek for sensitivity. The paper does not clearly report every graph size. Persistent and community-selected adversaries increased polarization and extremization. Increasing the manipulation budget strengthened the effect. Mitigations reduced the attack but did not restore baseline. Qualitative trends persisted across models and graphs.

Networks are fixed and opinions are reduced to one scalar. There is no validation against real social dynamics. Graph sizes and details are incomplete. Uncertainty is presented mainly through figures. The adversary is non-adaptive and scale is limited by cost. It does not demonstrate behavior of human populations. It does not estimate effects on a real platform. It does not show that any mitigation guarantees baseline recovery.

Español

Agentes con opinión numérica, terquedad y actividad generan posts y actualizan posición sobre grafos sintéticos HRG y grafos Twitter/Reddit. Se comparan selección aleatoria, grado, centralidad y comunidades; adversarios persistentes o susceptibles; moderadores reactivos e intervenciones de exposición, feed, filtrado, actividad y conexiones. Se promedian 10 ejecuciones.

Redes sintéticas y grafos derivados de Twitter y Reddit, con GPT-4.1-mini como modelo principal y GPT-4o-mini y DeepSeek en sensibilidad. El artículo no informa con claridad todos los tamaños de grafo. Los adversarios persistentes y seleccionados por comunidad aumentaron polarización y extremización. Aumentar el presupuesto de manipulación reforzó el efecto. Las mitigaciones redujeron el ataque pero no restauraron el baseline. Las tendencias cualitativas persistieron entre modelos y grafos.

Las redes son fijas y las opiniones se reducen a un escalar. No hay validación contra dinámica social real. Los tamaños y detalles de grafos no están completos. La incertidumbre se muestra principalmente mediante figuras. El adversario no se adapta y la escala está limitada por coste. No demuestra comportamiento de poblaciones humanas. No estima efectos reales de una plataforma. No prueba que una mitigación garantice recuperar el estado basal.

Research question

How much can a budget-limited adversary increase polarization in an LLM-agent network, and which mitigations reduce that effect?

Method

Agents with numeric opinion, stubbornness, and activity generate posts and update positions on synthetic HRG and Twitter/Reddit graphs. Random, degree, centrality, and community selection are compared; adversaries are persistent or susceptible; reactive moderators and exposure, feed, filtering, activity, and connection interventions are tested. Results average 10 runs.

Sample: Synthetic networks and graphs derived from Twitter and Reddit, with GPT-4.1-mini as the main model and GPT-4o-mini and DeepSeek for sensitivity. The paper does not clearly report every graph size.

Findings

  • Persistent and community-selected adversaries increased polarization and extremization.
  • Increasing the manipulation budget strengthened the effect.
  • Mitigations reduced the attack but did not restore baseline.
  • Qualitative trends persisted across models and graphs.

Limitations

  • Networks are fixed and opinions are reduced to one scalar.
  • There is no validation against real social dynamics.
  • Graph sizes and details are incomplete.
  • Uncertainty is presented mainly through figures.
  • The adversary is non-adaptive and scale is limited by cost.

What the study does not establish

  • It does not demonstrate behavior of human populations.
  • It does not estimate effects on a real platform.
  • It does not show that any mitigation guarantees baseline recovery.

Traceability

Scope: Full text

Version: arxiv; 14-page full text reviewed 2026-07-18

Consulted source: https://arxiv.org/abs/2606.18795

Review: Codex full-text and visual 14-page methodological, statistical and claim-boundary review, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4.1-mini
  • GPT-4o-mini
  • DeepSeek

Instruments and metrics

  • LLM social-network simulator
  • Polarization and extremization metrics

Data used

  • Synthetic HRG graphs
  • Twitter graph
  • Reddit graph

Evidence and location

  • Research question, method, results, and discussion: Full text, pp. 1-14, visually reviewed on 18/07/2026
  • Figures, tables, results, and limitations: Primary PDF sha256 025722221ef281a0e21e62d1769fd8a75f902ec1052bcb7628c78bdf563af68b; methods, results, limitations, and appendices
  • Editorial decision and claim boundary: Critical record article-411, complete cross-check of 14 pages