social influence modeling

This study systematically evaluates the efficacy and underlying mechanisms of AI-driven influence operations on social networks. By constructing a synthetic social network platform that integrates multi-agent simulation, natural language generation, and belief dynamics modeling, the work quantifies for the first time the impact of three core strategies—narrative dissemination, information amplification, and counter-messaging—on audience beliefs. The findings reveal that information amplification achieves the broadest reach, counter-messaging is most effective in shifting opinions, and narrative dissemination requires substantially greater adversarial investment to yield measurable effects. Furthermore, the research uncovers an intrinsic relationship between the behavioral footprints of influence actors and the resulting effectiveness of their campaigns.

This work addresses the challenge of modeling multi-stakeholder influence competition and countering misinformation in social networks. Methodologically, it introduces the first simulation framework that deeply integrates classical opinion dynamics (e.g., the DeGroot model) with multi-role large language model agents (Llama-3, GPT-4), augmented by graph neural networks to capture dynamic social topology—enabling interpretable, intervenable, and low-barrier simulation of social influence processes. Its contributions are threefold: (1) an open-source simulation platform empowering social scientists to conduct influence diffusion and intervention experiments without programming; (2) high-fidelity replication of information cascades, opinion polarization, and counter-misinformation strategy efficacy, validated on both real-world Twitter/X and synthetic social graph datasets; and (3) a novel paradigm for computational social science that bridges rigorous theoretical foundations with LLM-driven realism.

This study investigates whether large language model (LLM)-driven multi-agent systems can replicate core human social dynamics—conformity, group polarization, and community fragmentation—observed in online forums. Method: We employ a structured multi-agent simulation framework to systematically evaluate LLMs of varying parameter scales and reasoning capabilities (e.g., chain-of-thought prompting, self-consistency) on social influence tasks. Contribution/Results: We find that smaller-scale models exhibit stronger conformity under peer influence, whereas reasoning-optimized models demonstrate significantly greater belief stability and resistance to polarization. Critically, we provide the first quantitative evidence of a negative correlation between LLM cognitive capacity and susceptibility to social influence effects. These findings establish a reproducible methodological foundation and empirical basis for AI-augmented, controlled experiments in computational social science.

This study addresses the challenge of insufficient fine-grained behavioral representation and interpretable attribution in rumor propagation modeling. Methodologically, it introduces the first LLM-driven, configurable social agent simulation framework: it employs role-aware, behavior-controllable large language model agents to simulate rumor diffusion across four canonical synthetic social network topologies, integrating structural generation, dynamic propagation evaluation, and attribution analysis protocols. Key contributions include: (1) high-fidelity simulation at scale—supporting networks of up to 100 nodes and 1,000 edges; (2) precise control over diffusion coverage (0%–83%); and (3) the first systematic empirical validation of the synergistic impact of network structural heterogeneity and user role specification on rumor propagation pathways and reach. The framework establishes a reproducible, intervention-enabled simulation paradigm for computational social science and information ecosystem governance.

This study addresses three key challenges in modeling public opinion dynamics on online social platforms: (1) insufficient understanding of opinion evolution mechanisms, (2) scarcity of authentic influence data, and (3) ethical constraints inherent in traditional empirical analysis methods. To overcome these, we propose a novel simulation framework integrating large language models (LLMs) with agent-based modeling (ABM). Within this framework, agents autonomously generate content, evolve opinions, and dynamically reconfigure their attention networks. Crucially, we introduce proximal policy optimization (PPO)—a deep reinforcement learning algorithm—into the modeling of emergent opinion leadership for the first time. We further design a constrained action space and a self-observation mechanism to ensure policy stability and interpretability. Experimental results demonstrate robust, spontaneous emergence of opinion leaders across diverse scenarios. Learning curves confirm that the model autonomously discovers optimal influence strategies, exhibiting both adaptability and convergence in uncertain, dynamic information environments.

This study investigates how individual predictive capabilities and collaborative mechanisms jointly shape dynamic influence and enhance collective performance in multi-agent large language model negotiation. We propose the first formulation of the negotiation process as an input-dependent mixture-of-experts system grounded in the Friedkin–Johnsen opinion dynamics framework, demonstrating that an agent’s latent competence governs its influence. To make this competence observable, we introduce confidence level and alignment with initial opinions as proxy variables. Experimental results show that when the routing mechanism accurately reflects agents’ true capabilities, the system significantly outperforms both single-agent baselines and static ensemble approaches, thereby validating the effectiveness of our influence modeling and dynamic routing strategy.

Existing multi-agent simulations often overlook how differences in underlying models shape interaction dynamics. This study addresses this gap by constructing heterogeneous large language model (LLM) social networks, employing large-scale multi-agent simulation, content mediation analysis, and lexical pattern prediction. Results indicate that base models, rather than role assignments, primarily govern agent engagement, an effect that amplifies as network size increases. Furthermore, network dynamics under heterogeneous model compositions ultimately converge toward base model effects. By challenging the prevalent single-model assumption, this work demonstrates the cross-context predictability of base models and their decisive influence on engagement styles. These findings underscore the necessity of accounting for model heterogeneity in multi-agent research to better understand emergent social behaviors in LLM-driven simulations.

Existing social network models struggle to capture belief aggregation in populations of large language model (LLM) agents, as they neglect information filtering driven by limited attention and fail to distinguish genuine consensus from conformity. This work proposes the SNLA framework, which, for the first time, integrates agent attention mechanisms into social influence modeling by characterizing actual influence exerted rather than relying solely on network structure. We reveal how attention breadth and network topology jointly govern collective belief dynamics. Theoretically, we prove that narrow attention induces herding under bounded effective sample sizes, whereas wide attention recovers the wisdom of crowds only in undirected degree-regular graphs. Combining graph-theoretic analysis, tractable agent-based models, and multi-agent LLM simulations, we empirically validate a phase transition from herding to collective intelligence across controlled experiments and operator-controlled variants of three benchmark tasks.

This study establishes a microfoundational link between cross-sectional network models and their underlying generative behavioral mechanisms. Building on a continuous-time stochastic choice framework, it proposes a general modeling approach that accommodates non-network actors and multilateral relational constraints, yielding an exponential-family representation under equilibrium conditions to facilitate individual preference estimation. The key innovation lies in the natural decomposition of graph potential into a preference component reflecting agent utilities and an entropy component encoding tie-formation rules, thereby unifying behavioral and statistical network modeling paradigms. The framework is successfully applied to analyze friendship networks in professional organizations and to model structural phase transitions in small groups, demonstrating its empirical validity and broad applicability.

This work addresses the problem of fairly evaluating the individual contributions of seed users in social networks prior to information diffusion, a critical need for budget allocation, influencer pricing, and credit assignment under privacy constraints. Building upon cooperative game theory, we propose the first ex ante influence attribution framework based on the Shapley value. Theoretical analysis reveals that exact computation is feasible in polynomial time for single-step activation models, whereas the multi-step diffusion setting is #P-hard. To tackle this intractability, we develop an efficient approximation algorithm with formal theoretical guarantees. Extensive experiments on both real-world and synthetic networks demonstrate the effectiveness and scalability of our approach, offering a practical and equitable solution for ex ante influence attribution.