Year: 2025
Type: conference-paper
Abstract: This work presents an approach for designing AI agents that simulate social media activity by replacing Twitter conversations with large language models (LLMs).Using a time-series dataset of Twitter discussions about technologies (April 2019 -April 2020), we propose an approach that combines fine-tuned language models with timeline manager to capture both conversational dynamics and temporal posting patterns.This approach consists of two main components: 1) a timeline manager, which models posting frequency, reply behaviour, and temporal rhythms of users, and 2) conversation agents, fine-tuned for posting and replying within threads.We evaluate the system along two dimensions: structural accuracy (whether the timeline manager replicates conversation patterns and thread structures), and emotion dynamics (weather the emotion of synthetic data replicates the true emotion trends in the original dataset).Our results demonstrate that the proposed agent-based simulation captures key characteristics of real Twitter interactions, providing a foundation for large-scale synthetic social media ecosystems useful for studying information flow, emotion propagation, and the impact of emerging technologies.
Authors: Abdul Sittar, Mateja Smiljanić, Alenka Guček
Institution: Jožef Stefan Institute
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: gold
OpenAlex ID: W7225720360
