An Agent‑Based Simulation of Politicized Topics Using Large Language Models: Algorithmic Personalization and Polarization on Social Media

Abstract Digital platforms now act as the primary environments for public discourse, where recommender systems shape visibility, emotion, and interpretation. This study introduces the Recommender Systems LLMs Playground (RecSysLLMsP), a simulation framework designed to examine how algorithmic personalization interacts with language generation to influence engagement and polarization. The research provides a reproducible and transparent environment for testing algorithmic effects on collective reasoning, which is an issue central to democratic communication. The study employs a one‑hundred‑agent simulation grounded in psychometric and demographic data from Serbian social media users. Agents interact through five stages of progressively personalized content feeds mediated by LLM‑generated posts. Quantitative metrics such as engagement intensity, network modularity, sentiment variance and qualitative linguistic validation are used to assess behavioral and structural change. Results reveal that moderate personalization maximizes engagement, while full personalization reduces diversity and amplifies both structural and affective polarization (Q = 0.22 → 0.68). LLM‑based agents successfully reproduce realistic patterns of emotional contagion and ideological clustering. The implications extend to computational social science and policy. Simulation‑based experimentation can inform ethical recommender design and algorithmic governance. Limitations concern the absence of genuine human cognition. Thus, findings indicate systemic tendencies rather than behavioral prediction. Future research should integrate real‑world datasets, multilingual testing, and policy‑driven intervention modeling to further calibrate this digital “laboratory” for exploring AI‑mediated communication.