We introduce twony, a micro-simulation prototype designed to show the impact of online social network (OSN) mechanics - particularly recommendation algorithms - on emotional contagion and discourse dynamics. By harnessing the capabilities of Large Language Models (LLMs), the system generates an ecosystem of synthetic, politically engaged digital personas that interact within a controlled social media environment. These autonomous agents emulate human behaviors through in-context prompting techniques, enabling users to observe emotional transmission patterns under systematically varied conditions while circumventing the constraints inherent to real-user experimentation. The prototype implements two distinct recommendation paradigms: a baseline chronological feed and an emotion-prioritizing ranking mechanism that amplifies content based on emotional intensity, allowing the examination of the formation and reinforcement of echo chambers. Emotional valence is quantified via a fine-tuned BERT model, while network-level and agent-level visualizations track emotional cascades and polarization dynamics throughout the lifecycle. The system is architected as a browser-based application leveraging modern web technologies and decentralized APIs. Twony emphasizes accessibility, customizability, and extensibility. This contribution advances the field by providing a scalable, open-source prototype for systematically investigating OSN dynamics, ofering actionable insights for platform designers and policymakers seeking to mitigate harmful emotional contagion while fostering healthier online discourse environments.
The proliferation of OSNs has significantly transformed the nature of digital discourse [
CEUR
ceur-ws.org twony provides insights for researchers, policymakers, and platform designers seeking to mitigate the adverse efects of emotional amplification while fostering healthier online discourse environments.
Our prototype is designed to deliver a fully autonomous generation of synthetic social media feeds by leveraging advanced LLMs to create and emulate a wide array of digital personas. These personas interact dynamically within a simulated social media environment, mimicking real-world online behaviors and discussions. While the system is capable of initiating and running simulations without any manual input, it also ofers users the flexibility to intervene by manually posting content into the feed. This feature allows users to steer the direction of discussions toward specific topics of interest, enabling a more tailored and interactive experience. In the sections that follow, we provide a detailed overview of the specifications governing our agents, the architecture of the recommendation systems, and the methodologies employed for evaluation.
With recent advancements in generative AI and agent-based modeling, the key feature of our prototype
is the adaptation of LLMs to specific personalities. Our system provides access to three state-of-the-art
models: Llama 3.1 8B/70B [
During the simulations, the selected model is adapted to diferent personas, authentically emulating
individual social media users. We perform this alignment via in-context prompting, creating nuanced
digital identities [
Our interaction mechanics employ a structured approach based on predetermined rules and probabilistic action models. The system evaluates each agent’s likelihood of posting original content or replying to existing content at any given step. For replies, the system restricts candidate selections to the top- ranked posts and implements a randomized selection process within this filtered pool. Additional constraints prevent agents from responding to their own uncommented posts, commenting on their immediately preceding comments, or creating consecutive original posts. With these constraints, we aim to maintain natural conversation flow and prevent artificial interaction patterns.
We implement a fully connected network with a global shared feed. All agents receive the same ranked content which is displayed to users. The global feed can be sorted according to two distinct recommendation systems. As a baseline, we implement a chronological feed where the newest content appears at the top regardless of engagement metrics or emotional characteristics. Additionally, we define an emotion-based ranking that relies on the classification described in Section 2.3. The aggregated values — negative and positive valence — are combined to determine post-ranking. In the default settings, higher emotional intensity, regardless of valence type, yields a higher ranking. However, the system allows users to adjust the impact of these values within a range of ±1 (Fig. 2). Consequently, both valence types can be configured to have either a negative impact on ranking or be neutralized entirely. With the emotion-based ranking system, we aim to model an echo-chamber efect for emotional content, hypothesizing that emotional intensity increases over time as agents are exposed to progressively more emotional content [13].
For the evaluation, we utilize a pre-trained BERT classifier [
During the time of writing, the application is openly accessible at simon-muenker.github.io/ TWONy-micro/, with the source code available in the GitHub repository github.com/simon-muenker/ TWONy-micro. We leverage cutting-edge web technologies and decentralized architectures, emphasizing a high degree of customizability for end users and potential further adaptations.
Our prototype interface is implemented as a reactive JavaScript-based app that runs interactively as a browser application. The main system is served statically to the client and does not run server-side code. We utilize Astro as the website build engine in combination with Svelte [16] as the reactive UI framework. For designing our interface elements, we use the utility-first CSS framework Tailwind CSS [17]. For handling the local application state, we opt for the framework-agnostic store management system Nano Stores. We connect the LLMs and evaluation services via external APIs hosted separately on-premise. We deploy the LLMs via an Ollama backend through a customized API implemented in Python using FastAPI, exposing the necessary routes via a reversed NGINX proxy. The evaluation services are implemented in Python using the HuggingFace Transformers library [18] and FastAPI, also exposed through a reversed NGINX proxy. Through this decoupled architecture, we ensure streamlined adaptation and customization, such as replacing the provided LLM service with OpenAI or Anthropic interfaces.
The prototype demonstrates how recommendation algorithms that prioritize emotional content can potentially contribute to emotional contagion across digital platforms. By implementing both a neutral chronological feed and an emotion-prioritizing ranking mechanism, twony reveals how even simple algorithmic changes can significantly alter discourse patterns and emotional trajectories over time. Further, the simulation provides a controlled environment to observe the formation and reinforcement of echo chambers. By tracking emotional polarization at both network and agent levels, twony illuminates how recommendation mechanics can inadvertently create feedback loops that amplify emotional intensity. The LLM-based agent architecture ofers a novel perspective on user behavior modeling that bridges the gap between oversimplified theoretical models and ethically complex real-user experiments. By leveraging the sophisticated capabilities of LLMs to emulate human-like behaviors with consistent personas, the prototype provides a more nuanced representation of how diverse individuals might respond to and contribute to the emotional climate of online spaces.
The current prototype employs a fully connected network with a global shared feed, which simplifies the complex topological structures observed in real OSNs. Actual social platforms feature clustered communities, varied connection strengths, and asymmetric influence patterns that significantly impact information flow [ 19] and emotional contagion. This simplified network topology may not adequately capture the nuanced dynamics of community formation and inter-group interactions that characterize real-world platforms. While the prototype LLM-based agents ofer superficially convincing emulation of human-like behaviors, they remain imperfect approximations of actual user behavior. The personas, though derived from real social media data, cannot replicate the psychological complexity, contextual awareness, and historical experiences that shape human responses to social media content. Additionally, the current implementation lacks direct validation against observed human behaviors in comparable conditions. Also, our current focus on politically engaged users — who represent only a subset of typical social media users — creates an artificially engaged environment that doesn’t reflect the true heterogeneity of OSNs. Most users are passive consumers or engage only occasionally with political content [20], creating diferent network dynamics than our simulation currently represents. Further, The emotion classification system, while functional, employs a simplified model of human emotions. The reduction to positive and negative valence categories, while methodologically sound, may obscure more nuanced emotional responses that influence discourse dynamics. Furthermore, emotional contagion in humans involves complex psychological mechanisms that may not be fully captured by the current simulation. Also, the prototype’s current implementation does not account for external factors that significantly influence online discourse, such as breaking news events, seasonal trends, or platformspecific features like hashtags or groups. These contextual elements often serve as catalysts for emotional cascades in real OSNs and their absence may result in artificially stable or predictable simulation outcomes.
Our prototype ofers several promising applications across academic, and policy domains: Policy Maker Lobbying Policymakers and regulators can leverage the prototype to assess in a simpliifed manner potential impacts of recommendation systems on digital platforms to the emotionality of political discourse. By adjusting parameters, twony allows for the visualization of how abstract changes might influence discourse patterns and emotional dynamics.
Digital Literacy Education The visual and interactive nature of the prototype makes it a suitable educational tool for demonstrating how recommendation systems influence information consumption and emotional responses. Educational institutions can incorporate the prototype in media literacy curricula to help students understand the mechanics behind their social media experiences and develop more critical engagement with digital platforms.
Computational Social Science Research While technically limited and designed as a demonstration tool, the prototype could provide researchers with a controlled environment to systematically study emotional contagion efects, polarization dynamics, and information difusion patterns. This controlled testbed allows for isolating specific variables that would be dificult to manipulate in studies with real users, potentially advancing theoretical understanding of online social dynamics. With the inclusion of LLMs from diferent geographic origins (American, European, and Chinese), the system ofers an opportunity to examine how cultural contexts might influence online discourse patterns.
This work presents twony, a micro-simulation prototype designed to explore the impact of OSN mechanics on emotional discourse. By leveraging LLMs, twony simulates politically engaged digital personas interacting in a controlled social media environment. The prototype ofers a systematic framework to analyze emotional contagion efects and the role of recommendation algorithms in shaping discourse. Additionally, prototype implements two ranking mechanism that contrasts chronological feeds with emotion-prioritizing ranking, allowing for a demonstration of how diferent content ranking paradigms influence the spread and polarization of emotional content. The system’s open-source nature and modularity ensure extensibility for further research and policy evaluation, making it a valuable tool for studying digital discourse dynamics.
The prototype provides several insights into OSN dynamics. Most significantly, it showcases that recommendation algorithms that prioritize emotionally intense content can amplify emotional contagion and contribute to discourse polarization. It highlights how algorithmic curation plays a fundamental role in determining the tone and trajectory of online discussions. The prioritization of emotional intensity over time fosters self-reinforcing feedback loops, leading to the formation of echo chambers where exposure to diverse perspectives is minimized. This efect underscores the potential of digital platforms to shape ideological divides unintentionally. Furthermore, the use of LLM-based personas in twony exemplifies that agent-based simulation could ofer a viable alternative to real-user studies, providing a controlled and ethical environment for studying digital discourse.
While twony provides a first demonstration of emotional contagion in OSNs, several improvements
and expansions can enhance its expressiveness. One key area for improvement is the selection of agent
responses and post interactions. Future iterations should replace predefined interaction rules with
more dynamic selection models that leverage LLM-driven decision-making or statistical approaches
inspired by real-world behavioral data. Additionally, the current implementation of the prototype
supports only open-weight LLMs available through Ollama. Expanding this selection to include
closedweight models such as GPT-4 [21], Claude, or Gemini [22] through API integration would improve
the generative quality and enable broader comparative studies of digital discourse across diferent AI
paradigms. Another critical enhancement involves refining the network topology. The current version
of twony operates on a fully connected network, which does not accurately reflect the complexities of
real-world OSN structures. Introducing community formations, asymmetric influence patterns, and
varying connection strengths would better capture the nuances of online discourse [19]. Moreover,
improving the emotional classification system by moving beyond a simple positive/negative valence
model would provide a more accurate representation of emotional dynamics. Incorporating additional
emotional categories and enabling real-time adaptation to shifts in discourse tone could significantly
enhance the simulation’s realism. A significant next step should involve validating the realism of
agent behaviors before implementing more complex features. Drawing on methodologies from related
work [
We thank Kai Kugler and Nils Schwager for the constructive discussions. This work is fully supported by twon (project number 101095095), a research project funded by the European Union under the Horizon framework (HORIZON-CL2-2022-DEMOCRACY-01-07).
During the preparation of this work, the authors used Claude 3.7 and Grammarly in order to: Grammar and spelling check. After using these services, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [13] M. Del Vicario, G. Vivaldo, A. Bessi, F. Zollo, A. Scala, G. Caldarelli, W. Quattrociocchi, Echo chambers: Emotional contagion and group polarization on facebook, Scientific reports 6 (2016) 37825. [14] L. F. Barrett, J. A. Russell, The structure of current afect: Controversies and emerging consensus,
Current directions in psychological science 8 (1999) 10–14. [15] N. Chetty, S. Alathur, Trigger event and hate content: Insights from twitter analytics, in: 2019 International Conference on Advances in Computing and Communication Engineering (ICACCE), IEEE, 2019, pp. 1–5. [16] S. Bhardwaz, R. Godha, Svelte. js: The most loved framework today, in: 2023 2nd International
Conference for Innovation in Technology (INOCON), IEEE, 2023, pp. 1–7. [17] M. C. Klimm, Design Systems for Micro Frontends-An Investigation into the Development of Framework-Agnostic Design Systems using Svelte and Tailwind CSS, Ph.D. thesis, Hochschulbibliothek der Technischen Hochschule Köln, 2021. [18] T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, et al., Transformers: State-of-the-art natural language processing, in: Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, 2020, pp. 38–45. [19] P. Doreian, N. Conti, Social context, spatial structure and social network structure, Social networks 34 (2012) 32–46. [20] W. Gong, E.-P. Lim, F. Zhu, Characterizing silent users in social media communities, in: Proceedings of the International AAAI Conference on Web and Social Media, volume 9, 2015, pp. 140–149. [21] J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt,
S. Altman, S. Anadkat, et al., Gpt-4 technical report, arXiv preprint arXiv:2303.08774 (2023). [22] G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican, et al., Gemini: a family of highly capable multimodal models, arXiv preprint arXiv:2312.11805 (2023). [23] P. Törnberg, D. Valeeva, J. Uitermark, C. Bail, Simulating social media using large language models to evaluate alternative news feed algorithms, arXiv preprint arXiv:2310.05984 (2023).