Abstract
Understanding the dynamics of stance change on social media is crucial for addressing polarization and information integrity, yet observational studies face challenges including limited experimental control, restricted data access, and algorithmic confounds. We leverage Generative Agent-Based Modeling (GABM)—a novel simulation paradigm employing autonomous LLM-based agents to replicate human behavioral dynamics—to explore the predictors and mechanisms underlying stance change in a controlled, fully observable environment. We simulate a social media with 1,000 LLM-driven agents, equally split between Democratic and Republican profiles, engaged in discussions about the 2020 US election. Our analysis reveals that agents who change political orientation exhibit lower activity levels, reduced network centrality, and more polarized emotional expression compared to those maintaining consistent positions. Self-reported motivations cluster into four categories: desire for constructive conversation (47.9%), internal factors (20.8%), fact-checking influence (16.7%), and previous interactions (14.6%). While we do not claim agents replicate humans’ stance change behavior, the emergent patterns observed in our simulation qualitatively align with established findings from empirical research, suggesting that GABM may capture meaningful dynamics of opinion change.
1 Introduction
Opinion dynamics on social media emerge from the interplay of individual attitudes, interpersonal interactions, and structural properties of online networks [23]. Understanding how and why users adjust their stance in these environments is a central problem in computational social science, with implications for polarization, information integrity, and democratic participation [4, 20]. However, studying these dynamics in the real world poses substantial challenges: observational data offer limited control over confounding variables, social media platform APIs increasingly restrict data access, and privacy constraints impede fine-grained analysis of behavioral and affective signals. In addition, causal inference on stance change remains notoriously difficult due to the entanglement of algorithmic exposure [11].
Recent advances in Large Language Model (LLM)-driven generative agents offer a promising complementary methodology. The concept was first formalized by [18], who introduced agents endowed with LLM-powered memory, planning, and reflection capabilities. This led to the development of Generative Agent-Based Modeling (GABM), a novel simulation paradigm that employs autonomous LLM-based agents to model human behavioral dynamics through natural language understanding and contextual reasoning. In GABM simulations, generative agents act as theoretical probes: they enable controlled, interpretable, and fully observable experimental settings similar to artificial societies in traditional agent-based modeling [24]. Therefore, GABM offers a unique opportunity to isolate mechanisms, test hypotheses, and explore counterfactual conditions that are impossible or ethically impractical to study with human subjects. This perspective has motivated a growing body of work on social media simulation that investigates whether generative agents can reproduce social phenomena observed in real social media, from polarization and echo chambers [8, 14] to coordination patterns [15, 17].
Building on this line of research, we examine whether GABM simulations can be used to explore predictors and mechanisms underlying stance change. To our knowledge, this is the first study investigating the factors that drive generative agents to shift their positions during social media simulations. We instantiate this problem in the context of the 2020 US election, and conducted a GABM simulation with 1,000 generative agents, equally split between Democratic and Republican profiles. We define stance as political leaning (Democrat or Republican), and stance change as the transition from one political orientation to the other. Our objective is to characterize the behavioral, structural, and affective patterns associated with stance change, and to uncover the decision-making processes underlying these transitions. We operationalize this through two research questions (RQs):
What are the characteristics that distinguish agents who change stance versus those who do not?\
Which motivations drive stance change for generative agents?\
Our findings reveal that agents who change stance–hereafter, swingers–exhibit lower activity levels (posting and commenting less), reduced network centrality (lower PageRank, in-degree, and out-degree), and more polarized emotional expression compared to non-swingers. Analysis of self-reported motivations identified four primary drivers: desire for constructive conversation (47.9%), internal factors (20.8%), fact-checking influence (16.7%), and previous interactions (14.6%). Notably, several of these patterns align with prior work linking engagement levels [2, 5, 21, 22, 25], network centrality [7, 12], and openness to dialogue [1, 9, 13] to mechanisms of stance change in humans. We position this work as a first step in establishing a methodology where GABM simulations can generate testable hypotheses about the mechanisms driving opinion change. These simulations provide an environment where behavioral patterns, network effects, and affective signals can be isolated and systematically examined—conditions difficult to achieve with observational human data. Future work will directly compare these emergent patterns with empirical data from real users to assess the validity and limitations of generative agents as proxies for human behavior.
2 Materials and Methods
This section describes the data and simulation framework underlying our study. We first introduce the real-world Twitter dataset used to ground agent profiles in authentic political discourse, and then detail the GABM methodology adopted to simulate social media dynamics and capture stance change behavior.
2.1 Dataset
The dataset used in this study consists of tweets from the US 2020 Election dataset [6], a large-scale collection spanning June 2020 to December 2020. It contains more than 1.2 billion tweets generated by over 1 million unique users, capturing highly polarized discussions between liberal and conservative communities throughout the final months of the presidential campaign and its aftermath. This rich volume of interactions and user diversity provides a representative foundation for analyzing stance dynamics in online political discourse.
2.2 Methodology
We adopt the GABM framework introduced in [16] to simulate large-scale social media dynamics through LLM-driven generative agents. Each agent is modeled as an autonomous entity equipped with four core components. First, a personality profile is assigned based on real users tweets related to the 2020 US Election. Following [16], we use a prompt-based strategy with LLMs [19] to extract attributes such as ideological alignment, conversational tendencies, and engagement style (e.g., being outspoken, critical, or supportive of specific political figures). A pre-trained stance detection model [10] is then applied to each agent's extracted profile to establish its initial political orientation (Democratic or Republican).
Second, agents maintain a dynamic list of connections in a directed social media graph $ mathcal {G} = (mathcal {V}, mathcal {E})$, where nodes represent agents and edges encode follower–followee relationships.
Third, each agent maintains a memory module that stores previously generated posts and their received reactions. This memory is initialized empty to ensure that all behavioral dynamics emerge solely from in-simulation interactions rather than from prior external knowledge.
Fourth, a reasoning module powered by an LLM governs the agent's decision-making process. At each simulation iteration, agents choose among the set of available actions: (i) posting new content, (ii) commenting on others’ posts, (iii) liking content, (iv) re-sharing content, or (v) not interacting. This decision is influenced by the agent's personality profile, the feedback received from other agents during earlier simulation iterations (e.g., received comments, likes, or re-shares), and the content posted by other agents during the simulation, which is mediated by a recommender system. While the original framework supports preference-based recommendations [16], we adopt a simpler random recommendation strategy to isolate the behavioral mechanisms underlying stance change and avoid confounding effects from algorithmic filtering. Importantly, to guarantee robustness and internal consistency, the reasoning module incorporates an automatic re-prompting mechanism to handle potential LLM hallucinations, such as references to non-existent posts/users, thereby ensuring the reliability of agent behavior throughout the simulation.
After having decided their actions, agents enter an interview phase designed to elicit the motivations behind any political stance change. Agents whose stance diverges from the initial political orientation are prompted to explain the reasons for their position. The simulation proceeds for a fixed number of iterations, allowing agent behaviors and network dynamics to evolve over time.
3 Experiments
We simulated a social network with 1,000 generative agents, equally divided between Republican and Democratic profiles. Each agent is powered by Llama 3 70B1. The simulation runs for 10 iterations, during which agents publish new content and interact with others (e.g., commenting, liking, or re-sharing each other's content). All experiments were conducted on a system equipped with a 10th Gen Intel i9-10980XE processor, 256 GB of RAM, and an NVIDIA RTX A6000 with 48 GB memory. With this configuration, each iteration of the simulation takes approximately 4 hours to complete. We release our code to ensure reproducibility and support further research2. Statistical significance in our results is assessed using Mann-Whitney U test with FDR correction for multiple hypothesis testing.
3.1 RQ1: What are the characteristics that distinguish agents who change stance versus those who do not?
At the end of each simulation iteration, the content produced or engaged with by each agent (via re-sharing, commenting, or liking) was analyzed using the same stance detection model employed to establish their initial political leaning [10]. Based on these outputs, agents were categorized into two groups: we refer to as “swinger agents” those whose inferred stance diverges from their original political orientation, while “non-swinger agents” maintain consistent alignment throughout the simulation. Among the 1,000 agents, 142 exhibited stance shifts. To mitigate potential hallucinations of the LLM, we excluded cases where agents changed stance for only a single iteration before reverting to their original one. After this filtering, 96 agents were finally classified as swingers, with 58% (55 out of 96) shifting from a Republican to a Democratic stance, while the remaining 42% transitioned in the opposite direction. Subsequently, we conducted a comparative analysis between swinger and non-swinger agents across three complementary dimensions: (i) Activity, capturing the volume of actions performed and received during the simulation; (ii) Network Topology, assessing agents’ structural positions through connectivity measures; and (iii) Affective Dimensions, analyzing the sentiment and emotional tone expressed in their generated content.
3.1.1 Activity. The activity dimension assesses the frequency and type of actions performed and received by agents within the simulated environment (i.e., posting new content, not interacting, commenting, liking, or re-sharing). The comparison between the distributions of actions performed by swinger and non-swinger agents revealed statistically significant differences for two specific behaviors: posting new content (Figure 1a) and commenting (p < 0.05 for Mann–Whitney U test). In both cases, swinger agents exhibit lower activity levels than non-swinger agents, suggesting that stance change is associated with reduced content production. No significant differences were found for re-sharing and liking (p > 0.05). Furthermore, when analyzing the extent to which agents received interactions from others (i.e., received comments, likes, and re-shares), no statistically significant differences were observed between the two groups (all p > 0.05). This indicates that the likelihood of stance change is not influenced by the volume of external feedback received, but rather by the agents’ own active participation and conversational engagement within the simulated social environment.
3.1.2 Network Topology. Network topology analysis investigates whether agents’ structural positions and connectivity patterns within the simulated environment influence their likelihood of changing stance. We constructed a directed graph where edges represent interactions (i.e., comments, likes, or re-shares) from one agent to another. For each agent, we computed standard centrality metrics such as PageRank, betweenness centrality, in-degree, and out-degree to quantify their visibility and embedding within the network. The resulting distributions were compared between swinger and non-swinger agents using the Mann–Whitney U test. The analysis revealed statistically significant differences for PageRank, in-degree (Figure 1b), and out-degree (p < 0.05). In all three cases, swinger agents exhibit lower values than non-swinger agents, indicating that agents who are less central in the interaction network are more likely to reconsider their political stance. In contrast, betweenness centrality did not differ between the two groups (p > 0.05). To further explore relational exposure, we examined the political composition of each agent's follower network by calculating the proportion of Democratic and Republican agents they are connected to. No statistically significant differences emerged (p > 0.05), indicating that stance-changing agents are not embedded in more ideologically diverse neighborhoods than non-swinger agents. Combined with the lower PageRank, in-degree, and out-degree observed for swingers, this suggests that stance shifts are associated with reduced structural embeddedness, rather than with the diversity of social exposure.
3.1.3 Affective Dimensions. Affective analysis examines how the sentiment and emotional dynamics expressed by agents relate to stance change during the simulation. We first assessed sentiment using a transformer-based model3, which classifies content into positive, negative, or neutral categories [3]. The resulting sentiment distributions were compared between swinger and non-swinger agents using the Mann–Whitney U test. The analysis revealed statistically significant differences for both positive and negative sentiment (p < 0.05): swinger agents express significantly lower levels of positive sentiment (Figure 1c) and higher levels of negative sentiment compared to non-swinger agents. Neutral sentiment showed no significant variation between the two groups (p > 0.05). These findings indicate that stance-changing agents exhibit a more negative affective profile, suggesting that stance shifts are associated with a reduction in positive emotional tone. To capture more fine-grained emotional cues, we applied a second transformer-based model4 to detect primary emotions, including anger (Figure 1d), disgust, fear, joy, neutrality, sadness, and surprise. The comparison revealed statistically significant differences for all detected emotions (p < 0.05) except sadness.
Empirical grounding of RQ1 findings. The patterns observed in RQ1 align with prior findings in computational social science: lower content production among swingers is consistent with research linking active political participation to stronger ideological stability [2, 21, 22]; their reduced centrality reflects evidence that highly central users often act as opinion leaders while peripheral actors exhibit greater flexibility [7, 12]; and their more polarized emotional expression characterized by lower positive affect and higher intensity across most emotion categories resonates with prior studies [5, 25].
3.2 RQ2: Which motivations drive stance change for generative agents?
This research question explores the underlying motivations that drive stance changes among swinger agents. At the end of each simulation step, during the interview phase, agents who had shifted their stance were asked to articulate the reasons behind their change, providing short, discursive explanations. These textual explanations were manually analyzed to uncover recurring themes, resulting in four categories that capture the principal drivers of stance change. The most prevalent driver, expressed by nearly half of the swinger agents (47.9%), was a desire for constructive conversation. A second cluster (20.8%) grouped explanations related to internal factors, including emotional reflection, personal re-evaluation, or shifts in self-perception. The third category (16.7%) corresponded to fact-checking influence, encompassing references to verifying information or being exposed to credible counter-evidence. Finally, a smaller portion of agents (14.6%) attributed their stance change to previous interactions with other users. These explanations pointed to the impact of interpersonal exchanges and social reinforcement, suggesting that peer interactions—whether supportive or confrontational—can be a factor to determine stance shifts.
To better understand the relationship between these motivations and the stability of stance changes, we examined the temporal persistence of stance shifts. Specifically, we measure how many consecutive iterations each agent maintained a stance different from their initial political orientation. Figure 2 displays, for each motivation category, the distribution of agents across different persistence durations (from 1 to 9 consecutive iterations). We observe a clear asymmetry in the temporal patterns of stance change across motivational categories: while motivations such as fact-checking influence and previous interactions with other users tend to disappear quickly, agents driven by a desire for constructive conversation remain predominant in later iterations. This pattern suggests that agents motivated by openness to dialogue are more likely to maintain the newly adopted stance over time. A similar, though less pronounced, persistence is also observed for agents citing internal factors, further supporting the idea that more introspective or deliberative motivations correspond to longer-lasting stance transitions.
Empirical grounding of RQ2 findings. The motivations identified in RQ2 align with prior research showing that both a desire for constructive conversation (e.g., openness to opposing viewpoints) and fact-checking influence (e.g., exposure to minformation correction), can trigger belief updating [1, 9, 13].
4 Discussion
Our findings provide a tentative account of how behavioral, structural, and affective factors may interact in shaping stance change within the simulated environment. Swinger agents post and comment less frequently and occupy more peripheral positions in the interaction network. One possible interpretation is that highly active and central agents repeatedly articulate their views and receive social reinforcement, progressively consolidating their initial orientation. Less active and less central agents, by contrast, may experience weaker pressures toward ideological consistency and therefore remain more open to reconsidering their position. This interpretation is consistent with research linking political participation and network centrality to ideological stability and opinion leadership [2, 7, 12, 21, 22].
GABM therefore appears useful not only for reproducing patterns documented in empirical research [1, 5, 9, 13, 25], but also for connecting these patterns into candidate explanatory mechanisms. In this sense, GABM can serve as a hypothesis-generating testbed through which researchers and policymakers can formulate and examine testable accounts of opinion change and ideological stability under controlled conditions.
5 Conclusion and Future Work
This work examined whether LLM-driven generative agents can serve as theoretical probes for studying stance change in online environments. Using a simulation of 1,000 agents discussing the 2020 US Election, we analyzed which agents shift political orientation and why, drawing on behavioral logs, network structure, affective signals, and self-reported motivations. We found that swinger agents exhibit lower activity levels, reduced network centrality, and more polarized emotional expression (RQ1). Their explanations cluster around four themes, with constructive conversation (47.9%) and internal reflection (20.8%) emerging as dominant drivers that persist over time (RQ2).
Notably, several emergent patterns observed in our simulation are theoretically consistent with established findings on political participation, network centrality, and deliberative exposure. At the same time, these findings are limited on specific modeling choices, including the initial ideological distribution, the selected LLM architecture, the population size, the number of iterations, and the topical scope. Future work should systematically vary these parameters to assess the robustness and generalizability of the observed effects.
Beyond simulation-level validation, a critical next step is the systematic validation against human data. Additional extensions include introducing causal interventions (e.g., fact-checking injections, alternative recommendation mechanisms) and exploring different LLM architectures to better understand the conditions under which GABM can complement empirical research on political discourse.
Notes
Source
Imported from ACM’s structured HTML source. ACM Reference Format: Valerio La Gatta, Gian Marco Orlando, Marco Perillo, Ferdinando Tammaro, and Vincenzo Moscato. 2026. Who Changes Their Mind? Exploring Stance Shifts in LLM-Based Social Media Simulations. In 37th ACM Conference on Hypertext (HT '26), September 14--18, 2026, London, United Kingdom. ACM, New York, NY, USA 5 Pages. https://doi.org/10.1145/3800935.3830836
References
[1] Ingrid Bachmann and Sebastián Valenzuela. 2023. Studying the downstream effects of fact-checking on social media: Experiments on correction formats, belief accuracy, and media trust. Social Media+ Society 9, 2 (2023), 20563051231179694.
[2] Pablo Barberá. 2015. Birds of the same feather tweet together: Bayesian ideal point estimation using Twitter data. Political analysis 23, 1 (2015), 76–91.
[3] Francesco Barbieri, Jose Camacho-Collados, Luis Espinosa Anke, and Leonardo Neves. 2020. TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classification. In Findings of the Association for Computational Linguistics: EMNLP 2020. Association for Computational Linguistics, Online, 1644–1650. https://doi.org/10.18653/v1/2020.findings-emnlp.148
[4] Elisabetta Biondi, Chiara Boldrini, Andrea Passarella, and Marco Conti. 2023. Dynamics of Opinion Polarization. IEEE Transactions on Systems, Man, and Cybernetics: Systems 53, 9 (2023), 5381–5392. https://doi.org/10.1109/TSMC.2023.3268758
[5] Erica Cau, Virginia Morini, and Giulio Rossetti. 2024. Trends and topics: Characterizing echo chambers’ topological stability and in-group attitudes. PLOS Complex Systems 1, 2 (2024), e0000008.
[6] Emily Chen, Ashok Deb, and Emilio Ferrara. 2022. # Election2020: the first public Twitter dataset on the 2020 US Presidential election. Journal of Computational Social Science 5, 1 (2022), 1–18.
[7] Michael Conover, Jacob Ratkiewicz, Matthew Francisco, Bruno Gonçalves, Filippo Menczer, and Alessandro Flammini. 2011. Political polarization on twitter. In Proceedings of the international aaai conference on web and social media, Vol. 5. 89–96.
[8] Antonino Ferraro, Antonio Galli, Valerio La Gatta, Marco Postiglione, Gian Marco Orlando, Diego Russo, Giuseppe Riccio, Antonio Romano, and Vincenzo Moscato. 2024. Agent-based modelling meets generative ai in social network simulations. In International Conference on Advances in Social Networks Analysis and Mining. Springer, 155–170.
[9] Emma Hoes, Brian Aitken, Jingwen Zhang, Tomasz Gackowski, and Magdalena Wojcieszak. 2024. Prominent misinformation interventions reduce misperceptions but increase scepticism. Nature Human Behaviour 8, 8 (2024), 1545–1553.
[10] Kornraphop Kawintiranon and Lisa Singh. 2021. Knowledge Enhanced Masked Language Model for Stance Detection. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics. https://www.aclweb.org/anthology/2021.naacl-main.376
[11] Stephan Lewandowsky, Ronald E Robertson, and Renee DiResta. 2024. Challenges in understanding human-algorithm entanglement during online information consumption. Perspectives on Psychological Science 19, 5 (2024), 758–766.
[12] Patrick R Miller, Piotr S Bobkowski, Daniel Maliniak, and Ronald B Rapoport. 2015. Talking politics on Facebook: Network centrality and political discussion practices in social media. Political Research Quarterly 68, 2 (2015), 377–391.
[13] Diana C Mutz. 2002. Cross-cutting social networks: Testing democratic theory in practice. American Political Science Review 96, 1 (2002), 111–126.
[14] Masaya Ohagi. 2024. Polarization of Autonomous Generative AI Agents Under Echo Chambers. In Proceedings of the 14th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis. Association for Computational Linguistics, Bangkok, Thailand, 112–124. https://doi.org/10.18653/v1/2024.wassa-1.10
[15] Gian Marco Orlando, Valerio La Gatta, Diego Russo, and Vincenzo Moscato. 2025. Can Generative Agent-Based Modeling Replicate the Friendship Paradox in Social Media Simulations?. In Proceedings of the 17th ACM Web Science Conference 2025. 510–515.
[16] Gian Marco Orlando, Valerio La Gatta, Diego Russo, and Vincenzo Moscato. 2026. Validating generative agent-Based modeling in social media simulations through the lens of the friendship paradox. Information Processing & Management 63, 4 (2026), 104636.
[17] Gian Marco Orlando, Jinyi Ye, Valerio La Gatta, Mahdi Saeedi, Vincenzo Moscato, Emilio Ferrara, and Luca Luceri. 2025. Emergent Coordinated Behaviors in Networked LLM Agents: Modeling the Strategic Dynamics of Information Operations. arXiv preprint arXiv:2510.25003 (2025).
[18] Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology. 1–22.
[19] Heinrich Peters, Moran Cerf, and Sandra C. Matz. 2024. Large Language Models Can Infer Personality from Free-Form User Interactions. arxiv:2405.13052 [cs.HC] https://arxiv.org/abs/2405.13052
[20] Anton V. Proskurnikov, Alexey S. Matveev, and Ming Cao. 2016. Opinion Dynamics in Social Networks With Hostile Camps: Consensus vs. Polarization. IEEE Trans. Automat. Control 61, 6 (2016), 1524–1536. https://doi.org/10.1109/TAC.2015.2471655
[21] Michael D. Slater, Adam Shehata, and Jesper Strömbäck. [n. d.]. Reinforcing Spirals Model. John Wiley & Sons, Ltd, 1–11. https://doi.org/10.1002/9781119011071.iemp0134
[22] Chenhao Tan, Vlad Niculae, Cristian Danescu-Niculescu-Mizil, and Lillian Lee. 2016. Winning arguments: Interaction dynamics and persuasion strategies in good-faith online discussions. In Proceedings of the 25th international conference on world wide web. 613–624.
[23] Fei Xiong and Yun Liu. 2014. Opinion formation on social media: an empirical approach. Chaos: An Interdisciplinary Journal of Nonlinear Science 24, 1 (2014).
[24] Peijun Ye, Shuai Wang, and Fei-Yue Wang. 2018. A General Cognitive Architecture for Agent-Based Modeling in Artificial Societies. IEEE Transactions on Computational Social Systems 5, 1 (2018), 176–185. https://doi.org/10.1109/TCSS.2017.2777602
[25] Xiaojing Zhu, Cantay Caliskan, Dino P Christenson, Konstantinos Spiliopoulos, Dylan Walker, and Eric D Kolaczyk. 2023. Disentangling positive and negative partisanship in social media interactions using a coevolving latent space network with attractors model. Journal of the Royal Statistical Society Series A: Statistics in Society 186, 3 (2023), 463–480.
Do you like what you are reading? Subscribe to receive updates.
Unsubscribe anytime