Simulating Multi-Agent Social Media Ecosystem with LLMs: Modeling, Evaluation, and Exploratory Analysis of Ranking Mechanisms

Read full text Summarise with VIRAT
Authors: Abdul Sittar, Mateja Smiljanić, Alenka Guček, Achim Rettinger, Marko Grobelnik
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Simulating social media with large language models (LLMs) promises a controlled alternative to observational studies and live-platform experiments. However, existing LLM-based agent simulators rarely ground agent behaviour in real activity data or compare feed-ranking policies against empirical interaction networks. We present a data-driven multi-agent framework that combines a Temporal Fusion Transformer (TFT) scheduler predicting who acts, what they do, and on which target; retrieval-augmented generation (RAG) over historical posts; persona profiles; and LoRA fine-tuned LLaMA-2-7B generators, with pluggable chronological, engagement-based, and hybrid ranking. We evaluate the framework against real Reddit data in Technology, Climate Change, and COVID-19 at three scales (100 to 1,000 events). The TFT outperforms LSTM baselines on all next-event tasks; for example, macro-F1 for target-user prediction in Technology rises from 0.019 to 0.909. Replaying the real event sequence yields cascade-like structure and realistic interaction-type ratios in Technology but not in COVID-19 and Climate; because Technology also has the largest active population and the strongest retrieval, this cannot be attributed to scheduler quality alone. In single-run Technology comparisons, ranking is associated with differences in network topology: more even interaction distribution under chronological ranking, stronger hub concentration under engagement ranking, and stronger within-cluster, echo-chamber-like structure under hybrid ranking. Manual analysis of 368 interaction chains shows that semantic coherence weakens in long threads and that homophily is stylistic rather than topical.

Fields of Study & Subjects

Complex Network Analysis Techniques Misinformation and Its Impacts Opinion Dynamics and Social Influence

Publication details

Journal / source
IEEE Access
Pages
1–1
Publication date
1 January 2026
Language
EN