arXiv:2604.18011v1 [cs.SI] 20 Apr 2026
Topology-Aware LLM-Driven Social Simulation: A Unified Framework for Efficient and Realistic Agent Dynamics Yuwei Xu∗
Shulun Zhang∗
Yingli Zhou†
The Chinese University of Hong Kong, Shenzhen Shenzhen, China [email protected]
The Chinese University of Hong Kong, Shenzhen Shenzhen, China [email protected]
The Chinese University of Hong Kong, Shenzhen Shenzhen, China [email protected]
Shipei Zeng
Laks V.S. Lakshmanan
Chenhao Ma‡†
Shenzhen Research Institute of Big Data Shenzhen, China [email protected]
The University of British Columbia Vancouver, Canada [email protected]
The Chinese University of Hong Kong, Shenzhen Shenzhen, China [email protected]
ABSTRACT
1
Social simulation is essential for understanding collective human behavior by modeling how individual interactions give rise to largescale social dynamics. Recent advances in large language models (LLMs) have enabled multi-agent frameworks with human-like reasoning and communication capabilities. However, existing LLMbased simulations treat social networks as fixed communication scaffolds, failing to leverage the structural signals that shape behavioral convergence and heterogeneous influence in real-world systems, which often leads to inefficient and unrealistic dynamics. To address this challenge, we propose TopoSim, a unified topologyaware social simulation framework that explicitly integrates structural reasoning into agent interactions along two complementary dimensions. First, TopoSim aligns agents with similar structural roles and interaction contexts into shared backbone units, enabling coordinated updates that reduce redundant computation while preserving emergent social dynamics. Second, TopoSim models social influence as a structure-induced signal, introducing heterogeneous interaction patterns grounded in network topology rather than uniform influence assumptions. Extensive experiments across three social simulation frameworks and diverse datasets demonstrate that TopoSim achieves comparable or improved simulation fidelity while reducing token consumption by 50-90%. Moreover, our approach more accurately reproduces key structural phenomena observed in real-world social systems and exhibits strong generalization and scalability.
In real-world social systems, countless interactions among individuals occur every day, leading to complex and emergent macro-level social dynamics. Yet, directly observing and experimenting with such systems is often infeasible owing to their large scale and proprietary nature. To this end, social simulation addresses this challenge by enabling controlled experiments that model microlevel interactions and reveal macro-level social phenomena, which is crucial for broadening our understanding of complex social systems [2, 10, 11, 30]. In the literature, existing methods typically rely on classical statistical models and machine learning approaches [13, 16], which simulate social processes over fixed social structures using predefined or learned aggregate interaction rules. In practice, these approaches model state evolution by learning influence weights between individuals, applying threshold or averaging rules for opinion updates, or estimating probabilistic transition matrices that govern how states propagate across the population. However, complex social systems typically involve dynamically evolving interaction networks and rich individual attributes, making it difficult to design unified, fixed update rules, ultimately limiting the flexibility and generalizability of these methods. Recent advances in large language models (LLMs) have demonstrated strong reasoning and comprehension capabilities, enabling agents to mimic human-like cognition and communication [26]. Accordingly, LLM-driven social simulation has emerged as a promising paradigm for modeling complex social systems [29, 33, 35], in which each agent represents an individual human, enabling the simulation of large-scale social phenomena through iterative multiround interactions among multiple agents. At each interaction step, an agent gathers information from its neighbors, updates its internal state based on its own attributes and the gathered information, and then scatters newly generated messages back to its neighbors, as illustrated in Figure 1(a). However, we observe that such paradigms fail to fully utilize the structural effects inherent in real-world social systems, which play a critical role in shaping individual behavior as well as collective social dynamics. In real-world social networks such as X (formerly Twitter), individuals with similar structural
Reference Format: Yuwei Xu, Shulun Zhang, Yingli Zhou, Shipei Zeng, Laks V.S. Lakshmanan, and Chenhao Ma. Topology-Aware LLM-Driven Social Simulation: A Unified Framework for Efficient and Realistic Agent Dynamics. Artifact Availability: The source code, data, and/or other artifacts have been made available at https://github.com/D2I-CUHKSZ/MicroWorld.
∗ These authors contributed equally to this work. † Open-source project leader. ‡ Corresponding author.
INTRODUCTION
+10M Probably good.
❸ Scaer
❶ Gather
Uniform
Totally right.
❷ Update
Seems off.
(a) Simulation meanism
It’s perfect.
in existing simulation in reality +10M
Strong
Absolutely not. Not really.
(b) Aligned updates given similar contexts
Un ifo rm
Kinda true.
e l Li
(c) Source-agnostic v.s. source-aware
Figure 1: Agent behaviors and interaction patterns are coupled with network structure. roles tend to display convergent opinion dynamics over time due to shared interaction patterns (Figure 1(b)), while, during each individual’s state update, influence from neighbors is inherently heterogeneous, with high-profile public figures exerting substantially greater impact than ordinary users (Figure 1(c)). Although LLM-based agents can implicitly capture such structural effects through repeated interactions, this process is often slow and computationally expensive. This raises a fundamental question: How to extract and leverage structural signals from network topology to steer agent behaviors toward realistic social dynamics? Motivated by this view, we propose TopoSim, a topology-aware framework that explicitly incorporates structural signals from network topology to guide agent behavior and interaction dynamics. By making structure an explicit driver of simulation rather than a passive scaffold, TopoSim enables efficient and realistic LLM-based social simulation. This framework consists of two complementary components: an update coordination module that groups agents with similar update contexts, and a role-differentiation module that makes neighbor message exposure source-aware and topologyconditioned. Structural Signals for Update Coordination. Since LLMbased individuals update their states as a function of local interaction context, agents occupying similar structural roles are exposed to highly overlapping neighborhood information and tend to exhibit similar update patterns [23]. This implies that independently updating each agent’s state repeatedly processes highly overlapping contextual information, thereby incurring substantial redundancy in computation and token usage. By aligning updates according to topological roles, we coalesce agents with similar structural contexts into shared update units that undergo coordinated inference, substantially reducing computational cost while preserving emergent social dynamics. Structural Signals for Role Influence. In LLM-based social simulation, how individuals aggregate opinions from their neighbors is inherently shaped by network structure. As discussed earlier, existing paradigms often assume uniform influence across individuals, ignoring the asymmetric roles that pervade real-world social networks. In light of this, TopoSim models social influence as a structure-induced signal rather than a predefined attribute. By leveraging network topology to infer how individuals are differentially
influenced by their neighbors and how they exert influence over others, our framework directly induces heterogeneous roles from the network. Experimental Evaluation. We conduct extensive experiments across diverse networked social simulation settings to validate the generality and effectiveness of TopoSim. Specifically, we evaluate it on echo-chamber dynamics, social media opinion exchange, and news propagation. Results show that TopoSim matches the performance of full LLM-based simulations while reducing token usage by 50–90% on networks with up to 5000 individuals with 15k edges, whereas existing LLM-based social simulations are typically conducted on networks of only a few hundred nodes and thousands of edges. The study with real-world data further demonstrates that topology-driven role differentiation enables realistic and robust social dynamics across scenarios.
2 BACKGROUND 2.1 Graph-based Simulation From a graph-theoretic perspective, social simulation can be formulated as a message-driven dynamical system on graphs. A social network is represented as a graph 𝐺 = (𝑉 , 𝐸), where each node 𝑣𝑖 ∈ 𝑉 corresponds to an agent, and each edge (𝑣𝑖 , 𝑣 𝑗 ) ∈ 𝐸 denotes a channel of information flow from 𝑣𝑖 to 𝑣 𝑗 . Each agent maintains an internal state 𝑥𝑖𝑡 at time step 𝑡, encoding its opinion and behavioral tendency. The system evolves through iterative interactions. At each step, agent 𝑣𝑖 updates its state by aggregating information from its neighbors N (𝑣𝑖 ) = {𝑣 𝑗 | (𝑣𝑖 , 𝑣 𝑗 ) ∈ 𝐸}. The update rule is defined as 𝑥𝑖𝑡 +1 = U 𝑥𝑖𝑡 , {𝑥 𝑡𝑗 : 𝑣 𝑗 ∈ N (𝑣𝑖 )} , (1) where U is a general neighborhood-dependent update operator. This formulation provides a unified view that encompasses both classical opinion dynamics and agent-based simulation frameworks. Within this framework, different simulation paradigms primarily differ in how the operator U is instantiated. Classical models define U via hand-crafted rules or numerical update schemes [13, 16], whereas recent LLM-based simulators realize U through natural language interactions [35]. In LLM-based settings, each agent is initialized with a state 𝑥𝑖0 , which typically includes a profile (e.g., background or preferences) and an initial opinion that shapes its future updates. During simulation, agents interact by exchanging 2
3.1
12.0%
20.0% Others
8.0% 4.0% 0->-2 (n=7)
1->2 (n=6)
1->-1 (n=2)
-1->-2 (n=4)
-1->1 (n=2)
2->-2 (n=4)
1->2 (n=6)
2->-2 (n=3)
-2->-2 (n=6)
8.0%
14.0% 4.0%
12.0% 12.0%
6.0%
Figure 2: Consistency analysis of opinion update trajectories in ECS-50 simulation. free-form textual messages rather than structured numerical signals. These interactions are further organized through memory mechanisms: short-term memory captures recent exchanges, while long-term memory stores condensed summaries of historical experiences [29]. At each step, an agent updates its state by integrating newly received information with its memory, enabling a gradual and context-aware evolution of 𝑥𝑖𝑡 . The resulting trajectory can then be used to analyze macroscopic phenomena such as the evolution of public opinion.
2.2
Behavior Alignment
Graph-based social simulation treats the underlying network as a structural constraint on communication eligibility, driving message passing and opinion evolution that eventually gives rise to macro-level social phenomena. However, in LLM-based simulation, many agents exhibit highly similar behavioral trajectories despite being simulated individually. To illustrate this phenomenon, we run simulations on a 50-node graph under the EchoChamberSim [29] framework and group agents with similar opinion trajectories. As shown in Figure 2, the median and variance of these groups reveal that agents with similar initial and final opinions also tend to follow smooth and closely aligned evolution paths. The pie chart further shows that nine such trajectory groups account for about 80% of all agents in the simulation. A closer look at one representative group confirms that opinion updates are neither random nor highly volatile; instead, they evolve in a stable and largely consistent manner within the group. This observation suggests substantial computational redundancy when an LLM is applied to every individual agent, even when the goal is to reproduce macrolevel dynamics. Since many agents behave similarly, an effective mechanism for grouping structurally or behaviorally aligned agents could reduce simulation cost while preserving the emergent social patterns, which motivates the method we introduce next.
3
Topology-Aware Simulation Framework.
Motivated by the gap between network structure and simulation dynamics, we propose TopoSim, a topology-aware framework that treats graph topology not merely as a communication constraint, but as an active factor shaping agent interaction, as illustrated in Figure 3. Given an input social graph and the initial state of each agent, TopoSim updates the agent states through iterative LLMbased simulation, and ultimately yields the simulated evolution of opinions and behaviors. The core idea of TopoSim is realized through two complementary mechanisms. The first is update coordination, which reduces redundant reasoning by exploiting coherence of structural similarity and similar update among agents. The second is role differentiation, which restores asymmetric influence by leveraging topology-derived importance. Together, these two designs enable the simulator to be both more efficient and more faithful to the heterogeneity of real social systems. Update coordination: structureaware inference sharing. Agents occupying similar structural positions are often exposed to comparable neighborhood signals and therefore tend to undergo similar updates, especially when their current states are also aligned. Based on this observation, TopoSim groups agents with similar structural contexts and states into shared units. Instead of invoking the LLM separately for every agent, we perform inference once per unit and share the resulting update across all agents in that unit. This unit-level reasoning backbone preserves the open-ended and context-sensitive generation ability of LLMs, while avoiding repeated computation on near-duplicated inputs. Role differentiation: topology-induced asymmetric influence. In real networks, agents do not contribute equally to collective dynamics: central or well-connected individuals often exert disproportionately larger influence on others. To capture this asymmetry, TopoSim uses topology-derived importance to modulate how information from different agents affects subsequent updates. Such influence patterns are induced directly by the network structure, without manual rules or external supervision, allowing the simulation process to reflect heterogeneous social roles more faithfully. Together, these two mechanisms define a topology-aware update operator U, which serves as the foundation of our simulation framework. As a result, TopoSim supports LLM-driven social simulation that is both computationally efficient and structurally faithful. We describe the coordination and differentiation modules in Sections 3.2 and 3.3, respectively.
3.2
Topology-induced Update Coordination
In social network simulations, agents occupying similar structural positions would gather comparable neighborhood messages for update. As a result, even expressive and stochastic LLM-based update operators U often produce similar state transitions across such agents. This redundancy eliminates the need for independent inference on each agent. We therefore introduce coordination as a modeling principle, where agents with similar structural contexts are jointly updated, reducing redundant inference while preserving the expressive power of LLM. Coordination emerges as agents align their updates through shared message exposure and compatible internal states. Directly identifying such joint conditions across all node pairs, however, is
METHOD
In this section, we present TopoSim, a topology-aware framework for LLM-driven multi-agent social simulation. We first introduce the overall formulation of TopoSim in Section 3.1. Then we detail its two key components, update coordination and role differentiation, in Sections 3.2 and 3.3. Finally, we provide a property analysis in Section 3.4. 3
Figure 3: Coordinated LLM updates (top) and emergent role differentiation (bottom) in TopoSim. is a subset 𝐶 ⊆ 𝑉 where 𝜅𝑖 𝑗 ≥ 𝜏 for all 𝑣𝑖 , 𝑣 𝑗 ∈ 𝐶; agents with no such peer form singleton units. With coordination units constructed, simulation proceeds at the unit level. Each unit 𝐶 ∈ C is updated via a single LLM invocation that synthesizes a transition for all agents within 𝐶, homogenizing their responses to the aggregated interaction context. To enhance representativeness, we select a representative agent 𝑣𝑟𝐶 ∈ 𝐶 as: ∑︁ 1 ∑︁ 𝑣𝑟𝐶 = arg max ( 𝜅𝑖 𝑗 + 𝛽 · 𝜋𝑘𝑖 ), (6) 𝑣𝑖 ∈𝐶 |𝐶 | 𝑣 ∈𝐶 𝑣 ∈𝑉
computationally prohibitive as it involves evaluating all possible neighborhood influenced inputs. To address this, TopoSim leverages structural similarity as a structure-derived prior: agents occupying similar topological positions are more likely to receive comparable message flows [24]. We first construct a candidate set of agents via coarse topology-aware filtering in their structural embeddings 𝜇𝑖 and 𝜇 𝑗 : Ecand = { (𝑣𝑖 , 𝑣 𝑗 ) ∈ 𝑉 × 𝑉 | cos(𝜇𝑖 , 𝜇 𝑗 ) ≥ 𝛾 },
(2)
where 𝛾 is a structural similarity threshold. These candidates are then refined by explicitly measuring the consistency between their states and neighborhood messages. Only pairs exhibiting sufficient consistency would be grouped into units for coordinated update. This two-stage approach leverages structural signals to efficiently identify high-quality candidate pairs. Building on the structurally filtered candidates, we refine coordination by enforcing consistency in both neighborhood messages and internal states. To quantify neighborhood exposure similarity, we first construct a categorical opinion distribution for each node. Specifically, each opinion message 𝑥 𝑡𝑗 is mapped to a discrete opinion category 𝜙 (𝑥 𝑡𝑗 ) ∈ {𝑐 1, . . . , 𝑐 𝐾 } (e.g., strongly oppose to strongly support). Let e(𝑐𝑘 ) ∈ R𝐾 denote the one-hot encoding of category 𝑐𝑘 . Then the neighborhood opinion distribution of Í agent 𝑣𝑖 is defined as 𝑣˜𝑖𝑡 := 𝑗 ∈ N (𝑖 ) 𝜋 𝑖𝑗 e 𝜙 (𝑥 𝑡𝑗 ) , where 𝜋 𝑖𝑗 denotes the structural influence of 𝑣 𝑗 on 𝑣𝑖 . Neighborhood similarity can be formulated with the Jensen–Shannon divergence between the relevant neighbor opinion distributions: 𝑡 𝑡 𝑠𝑖info ∀(𝑣𝑖 , 𝑣 𝑗 ) ∈ Ecand . (3) 𝑗 = 1 − 𝐷 JS (𝑣˜𝑖 , 𝑣˜ 𝑗 ),
𝑗
3.3
Topology-aware Asymmetric Influence
In social networks, users exert different levels of influence on one another, even when exposed to similar neighborhood information. For instance, people tend to trust information from influencers or established news sources, yet remain more resistant to peers of similar status. Conversely, public figures rarely revise their views in response to ordinary fans. Crucially, these social roles emerge from the network topology, such as the number of followers. However, most LLM-driven simulators still assume homogeneous neighbor influence, treating all observed messages as equally informative. We therefore introduce role differentiation as a modeling principle, where topology determines how external information is weighted during agent reasoning, enabling structurally grounded asymmetric influence. From a social modeling perspective, network topology induces a latent influence distribution that governs how information is treated during reasoning. Rather than assigning agents fixed social roles, we model influence as a latent variable 𝜋 𝑖 defined on the neighborhood of agent 𝑣𝑖 , where 𝜋 𝑖𝑗 quantifies the expected contribution of
For internal state compatibility, we represent each agent’s state as a numerical feature vector z𝑖𝑡 = [ 𝜙 (𝑥𝑖𝑡 ), 𝑠𝑖 ] ⊤ , where 𝜙 (𝑥𝑖𝑡 ) is the same opinion score as above, and 𝑠𝑖 = S(𝑥𝑖𝑡 ) quantifies stubbornness of 𝑣𝑖 . State divergence is measured by Euclidean distance: 𝑠𝑖state = z𝑖𝑡 − z𝑡𝑗 , ∀(𝑣𝑖 , 𝑣 𝑗 ) ∈ Ecand . (4) 𝑗 We unify these signals into a continuous consistency score: state 𝜅𝑖 𝑗 = 𝑠𝑖info 𝑗 · exp(−𝜆𝑠𝑖 𝑗 ),
𝑘
where the first term measures the average consistency of 𝑣𝑖 with other members in the unit, and the second term quantifies its global influence. Their influence weight can be adjusted with 𝛽. The LLMbased update generated for 𝑣𝑟𝐶 is shared across all members of 𝐶, defining a coordinated transition. This reduces the number of LLM invocations from |𝑉 | to |C|, forming an efficient backbone for network-wide simulation while preserving semantic fidelity with minimal loss of behavioral variability.
(5)
where the exponential term softly penalizes state mismatch while preserving sensitivity to minor discrepancies. A coordination unit 4
where D denotes a distributional divergence, 𝑥 encodes the agent’s state, and 𝑣˜ represents its aggregated neighborhood opinion. This assumption states that, at the level of output distributions, the update produced by U changes smoothly with respect to two perturbation sources: the agent’s own state and the aggregated neighborhood opinion. The constants 𝐿𝑥 and 𝐿𝑣 quantify the sensitivity of the update operator to these two factors, respectively. In this way, if two agents receive similar state inputs and neighborhood opinions, their induced output distributions should remain close. For a coordination unit 𝐶, let 𝑣𝑟𝐶 denote its representative agent. The unit-construction criterion is designed to ensure that agents grouped into the same unit have similar internal states and similar neighborhood opinion compositions. We therefore summarize the within-unit discrepancy by two quantities: a state discrepancy 𝛿𝐶 and a neighborhood-distribution discrepancy 𝜖𝐶 . Once 𝑣𝑟𝐶 is used to replace the individual update of each 𝑣𝑖 ∈ 𝐶, the representativebased approximation follows directly from the assumption: D U (𝑥𝑖𝑡 , 𝑣˜𝑖𝑡 ), U (𝑥𝑟𝑡𝐶 , 𝑣˜𝑟𝑡𝐶 ) ≤ 𝐿𝑥 𝛿𝐶 + 𝐿𝑣 𝜖𝐶 . (10)
messages from 𝑣 𝑗 to the update of 𝑣𝑖 . This latent variable is not a role assignment, but a structurally induced weighting that is both asymmetric and target-specific. To instantiate 𝜋 𝑖 , we adopt Personalized PageRank (PPR), which arises as the stationary distribution of a random walk with restart at 𝑣𝑖 : 𝜋 𝑖 = (1 − 𝛼)𝑒𝑖 + 𝛼𝑊 𝜋 𝑖 , (7) where 𝑒𝑖 is the personalization vector and 𝑊 is the normalized transition matrix. Value 𝜋 𝑖𝑗 naturally quantifies the structural influence of 𝑣 𝑗 on 𝑣𝑖 , integrating proximity and path-based reachability. Interpreted this way, PPR serves as a topology-induced influence model rather than a centrality heuristic, providing a scalable, annotationfree mechanism for asymmetric social influence. Given the influence profile 𝜋 𝑖𝑗 , we incorporate asymmetry into agent reasoning by modulating how neighbor signals are exposed to the LLM. For agent 𝑣𝑖 , incoming messages are no longer treated uniformly, but aggregated as: 𝑚𝑖𝑡 = AGG 𝑗 ∈ N (𝑖 ) 𝜋 𝑖𝑗 , Φ(𝑥 𝑡𝑗 ) , (8) where Φ(·) extracts communicative text message from a neighbor’s state. To ensure that influence differences are preserved in the LLM, we do not directly inject numerical weights into the prompt. Instead, neighbors are automatically grouped into a small number of importance tiers derived from 𝜋 𝑖𝑗 , and presented to the LLM as prioritized signal groups (e.g., core vs. peripheral information), together with corresponding attention instructions. This aligns the prompt structure with the graph, allowing topology-induced asymmetry to directly shape the LLM’s inference process in a stable and interpretable manner. As a result, differentiated social roles and asymmetric influence emerge automatically from topology, without manual role design or supervision, while remaining fully scalable to large simulations.
3.4
This gives a simple upper bound on the distributional error induced by coordinated inference: the approximation error is controlled by how much the grouped agents differ in state and in neighborhood opinion distribution. Therefore, tighter coordination units lead to smaller approximation error, while larger units provide greater efficiency gains, yielding an explicit efficiency–fidelity trade-off under the assumption.
4
In this section, we present a comprehensive evaluation of TopoSim in LLM-driven social simulations. We first introduce the experimental setup. Then we aim to answer the fundamental question: can structural signals be leveraged to improve the dynamics of multiagent simulation? We evaluate it from four complementary aspects: 𝑄 1 whether update coordination reduces cost while preserving dynamics, 𝑄 2 whether role differentiation leads to more realistic patterns, 𝑄 3 whether their combination enhances realism and efficiency, 𝑄 4 whether these benefits are robust across different LLM backends, and 𝑄 5 how scalable is TopoSim.
Approximation Analysis of Coordination
Coordination accelerates LLM-driven simulation by replacing nodelevel updates with shared unit-level transitions, which inevitably introduces approximation error. Since U is implemented by a stochastic language model, a strict pointwise analysis is not appropriate. Instead, we present an assumption-based stability argument showing that the coordination error remains controlled when the grouped agents have similar inputs. Let 𝑥𝑖𝑡 denote the state of agent 𝑣𝑖 at step 𝑡, and let the original update be 𝑥𝑖𝑡 +1 = U (𝑥𝑖𝑡 , 𝑣ˆ𝑖𝑡 ), where 𝑣ˆ𝑖𝑡 is the gathered neighborhood messages. For analysis, we summarize neighborhood input by a Í discrete opinion distribution 𝑣˜𝑖𝑡 := 𝑗 ∈ N (𝑖 ) e 𝜙 (𝑥 𝑡𝑗 ) . This abstraction is natural in our simulation setting for two reasons. First, the prompts used for agent updates are highly templated, so the variation in LLM input is mainly driven by changes in agent state and the composition of observed neighbor opinions, rather than arbitrary prompt-form changes. Second, after each simulation step, the generated opinion is mapped back to a discrete opinion category, so the downstream dynamics depend primarily on coarse opinion shifts rather than fine-grained linguistic variation. Motivated by this setting, we assume that the LLM-driven update operator is Lipschitz continuous in distribution [4]: ˆ U (𝑥 ′, 𝑣ˆ′ )) ≤ 𝐿𝑥 |𝑥 − 𝑥 ′ | + 𝐿𝑣 𝐷 JS (𝑣, ˜ 𝑣˜ ′ ), D (U (𝑥, 𝑣),
EXPERIMENT
4.1
Experimental Setup
System setup. All experiments were conducted on an Ubuntu 24.04 server with two Intel Xeon Gold 6330 CPUs, 1 TB RAM, with one NVIDIA RTX A5000 GPU. The implementation was done using Python. Experimental frameworks. We evaluated TopoSim on three public LLM-based social simulation frameworks. EchoChamberSim (ECS) [29] offers a controlled testbed with standard opiniondynamics metrics and configurable topologies, enabling precise measurement of macro behaviors (e.g., polarization). OASIS [33] provides an end-to-end, platform-style social media simulator for testing whether structural signals improve realism under realistic system conditions. FDE-LLM [34] is grounded in real user discussions and targets reproducing observed group attitude trajectories over time, allowing direct assessment of alignment with real-world
(9) 5
Figure 4: Impact of Update Coordination on Social Dynamics Across ECS and OASIS.
# Nodes
# Edges
# Timesteps
Data Type
ECS-50 ECS-500 OASIS Weibo
50 500 196 206
108 1,753 1,279 1,503
30 10 80 18
Synthetic Synthetic Synthetic Real-world
0 (0.0,0.25]
Mean Absolute Difference
dynamics. Collectively, they test whether topology-aware coordination and role differentiation improve efficiency, realism, and robustness. Experimental datasets. We evaluate TopoSim on three social simulation datasets (Table 1). ECS is a synthetic network with LLM-generated agent profiles, designed to model echo chambers at configurable scales. OASIS [33] consists of 196 real X accounts with profiles and connections, where agents discuss a moral dilemma in conservative and progressive settings. Weibo [34] captures around 200 real users discussing a trending news event over three days, providing a benchmark for behavioral realism. Experimental settings. We use gpt-4o-mini as the backbone LLM for all social simulation experiments, following the default configuration of each framework and setting the temperature to 0.7. To test robustness across LLM backbones, we additionally report results with Qwen-plus and DeepSeek-V3.2. For topology-aware signals, we compute Personalized PageRank [19] on the underlying interaction graph with a damping factor of 0.85. When deriving structure-aware node representations via struc2vec [28], we set the embedding dimension to 128. Metrics. In our quantitative evaluation, we adopt the three standard metrics used in ECS: Polarization (Pol.), which measures the overall dispersion of opinion values; Global Disagreement (GD), which characterizes the average intensity of disagreement between each node and its neighbors; and Normalized Clustering Index (NCI), which measures the consistency between a node’s opinion and the mean opinion of its neighbors.
4.2
Variance bins
Epochs
(0.25,0.5] (0.5,0.75]
ECS-500 Proportion
Name
ECS-50 Proportion
Table 1: Dataset statistics.
(0.75,1] ＞1.0
Mean Absolute Difference
Epochs
Figure 5: Node-level consistency diagnostics on ECS. Figure 4 reports the evolution of macro-level metrics over simulation steps. Overall, the coordinated simulation remains highly similar to the baseline on all three metrics across both frameworks. Although the coordinated version shows slightly higher variability in the early steps, it quickly converges to stable dynamics. By the end of the simulation, the differences remain very small. For instance, in all four scenarios, polarization and global disagreement differ by less than 0.2, while NCI differs by less than 0.1. These results show that topology-aware coordination preserves the overall social dynamics with minimal deviation from full-agent simulation. 2.Intra-unit consistency. To further validate whether coordination can preserve simulation behavior, we examine node consistency within coordinated units by comparing with the full-agent simulated results on ECS, as shown in Figure 5. The left panel shows the mean absolute difference of the agent’s opinion between Base and Coordination simulation results. On both scales, the distribution is tightly concentrated near zero, indicating that most nodes maintain behavior highly similar to that in the full-agent simulation. The right panel analyzes the variance distributions of agents’ opinions within each coordinated unit under full-agent simulation, which shows the original diversity within units. The right panel presents the distribution of opinion variance within each coordinated unit when under full-agent simulation, reflecting opinion diversity among agents within each unit. In ECS-500, over 80% of units exhibit a variance below 0.25 — a threshold that can be interpreted as: given 15 agents in unit, it is violated only if one agent’s opinion deviates from the group mean by 2. The low variance indicates that agents assigned to the same unit tend to exhibit
Coordination for Efficiency
1.Effectiveness. To answer 𝑄 1 , we conduct experiments on two simulation frameworks, ECS and OASIS, comparing the full-agent simulation (Base) with our topology-aware coordination approach. 6
Figure 6: Topology-aware role differentiation on OASIS.
Figure 7: Role Differentiation Effect and Ablation Study on Real-world dataet.
similar opinion trajectories throughout the simulation, confirming that the coordination mechanism effectively groups nodes with similar opinion dynamics. 3.Efficiency. While preserving both macro-level dynamics and node-level consistency, topology-aware coordination substantially reduces token consumption, as summarized in Table 2. The savings increase with scale, reaching 77.3% on ECS-500, where the larger graph allows more low-impact peripheral agents to be coordinated, thereby avoiding redundant inference calls. The method is even more effective on OASIS, achieving 91.7% token savings, reducing total token usage from 4.53M to 376K. The substantial decrease results from denser interaction structure and richer contextual exchanges in OASIS, making coordination especially beneficial. Overall, these results demonstrate that topology-aware coordination offers a favorable trade-off between efficiency and effectiveness, significantly reducing simulation costs while preserving the essential social dynamics.
although most agents adopt progressive opinions, a noticeable minority shifts toward conservative views. To further validate this observation, we evaluate our approach on FDE-LLM, using data that captures real-world opinion shifts over time. Real Data. To further validate realism, we evaluate the method on FDE-LLM, which captures real-world opinion shifts over time. Figure 7(a) compares the mean opinion trajectories of the ground truth (GT), the baseline simulation, and the role-differentiated variant (RD), together with their trajectory similarity to GT over time. The baseline gradually deviates from the real trajectory, with similarity dropping to around 0.6, whereas RD remains closely aligned with the ground truth throughout simulation and maintains over 90% similarity throughout the simulation. This result indicates that role differentiation substantially improves the realism of simulated opinion evolution, enabling the model to better capture the finegrained temporal patterns observed in real-world social dynamics.
4.4
Table 2: Token Savings Across Frameworks.
4.3
Framework
Scale
Base Token
Co-Token
Token save
ECS
50
320,235
185,556
43.00%
FDE-LLM
206
286,533
136,736
52.23%
ECS
500
3,281,979
743,775
77.30%
OASIS
196
4,526,566
376,677
91.70%
Ablation Study
To better illustrate the roles of different components in our framework, we conduct an ablation study comparing four variants on Weibo dataset: Baseline, Coord., RD, and Hyb.. Figure 7 reports their similarity score over simulation steps and their token cost. Similarity. The line chart of Figure 7(b) shows the similarity to the ground-truth opinion trajectory over time. Compared with the baseline, both RD and Hyb. maintain consistently high similarity throughout the simulation, indicating that role differentiation is the main factor improving realism. In contrast, Coord. alone preserves reasonable alignment in early stages but becomes less stable at later timesteps. Cost. The bar chart of Figure 7(b) shows the token cost of different variants. Coord. achieves the lowest cost, confirming that update coordination is the primary source of efficiency gain. RD remains close to the baseline in cost, since it mainly changes the interaction mechanism rather than reducing the number of updates. Hyb. lies between Coord. and Baseline. Overall, the ablation results reveal a division of labor between the two topology-aware components: role differentiation primarily improves simulation realism, while the coordination mechanism
Role Differentiation for Realism
To answer 𝑄 2 and 𝑄 3 , we evaluate topology-aware role differentiation on two types of data sources: synthetic data in OASIS and real-world data in FDE-LLM. Synthetic Data. Figure 6(a) shows the OASIS results under two ideological configurations, reporting the proportions of agents with different opinions after simulation. The results indicate that incorporating role differentiation yields more nuanced opinion dynamics. Under conservative settings, a larger fraction of agents remains at Draw rather than becoming polarized. Under progressive settings, 7
Base
NCI
GD
major research direction models via statistical formulations, where local influence rules – such as Friedkin–Johnsen opinion updating and bounded-confidence interactions – are defined over a fixed network [7, 13, 21]. These formulations show that simple update mechanisms can give rise to global patterns such as consensus or polarization. Beyond analytical models, agent-based simulations view societies as collections of autonomous agents interacting on realistic network topologies, such as small-world graphs [3, 9, 31, 36]. Such models demonstrate that complex social structures, from cultural clusters to social hierarchies, can naturally emerge from agent behaviors. Together, these works establish a bottom-up paradigm for social simulation. Structural Influence. A long-standing principle in social network analysis holds that an individual’s behavior and influence are deeply shaped by their structural position. Nodes occupying equivalent roles tend to exhibit similar interaction patterns, regardless of their attributes [23, 32]. This insight inspired computational frameworks for role discovery, ranging from blockmodels [32] to representation learning approaches such as RolX [17] and struc2vec [28] that explicitly encode topological similarity into node embeddings. Concurrently, centrality measures like personalized PageRank (PPR) [19, 25] have been widely adopted to quantify asymmetric influence in social dynamics [1, 8], reflecting how network context modulates persuasiveness. LLM-driven Simulation. More recently, advances in LLMs are transforming social simulation, enabling intelligent and interpretable agents to be integrated into networked social environments. Early study demonstrated that LLMs could simulate human-like cognition, enabling lifelike agents in sandbox worlds [15]. This paradigm soon extended to social networks, where agents interacting over them reproduced emergent phenomena such as polarization and emotion propagation [5, 12, 14]. In addition to dynamic changes at the macro level, recent simulators can also capture micro-level changes in message propagation [6, 22]. Concurrently, the simulations also pursued scalability and realism. Simulators would support large-scale social experiments [18, 20, 27], while others model structured communities via heterogeneous graphs [35]. Despite extensive evidence that social networks shape behavior, the network structure in most LLM-based simulators remains a passive communication channel. We seek to fill this void in this work.
Pol. / GD
NCI
Pol.
Pol. / GD
Coordination
NCI
Iteration days
Iteration days
ECS-Qwen-Plus
ECS-Deepseek (a) Robustness
NCI
Pol. / GD
Iteration days
Iteration days
Base
Coordination (b) Scalability (ECS-5000)
Figure 8: Cross-LLM robustness and scalability on ECS. Table 3: ECS Token Savings Across Scales. Scale
Base Token
Co-Token
Token save
1000
12,238,903
3,794,304
69.00%
2000
26,225,807
7,821,154
70.18%
5000
77,873,764
17,637,028
77.35%
primarily improves efficiency. Their hybrid combination provides a practical trade-off between fidelity and cost.
4.5
Robustness and Scalability
To answer 𝑄 4 and 𝑄 5 , we conduct two additional studies: crossLLM evaluation and large-scale simulation, to further assess the robustness and scalability of our approach. Cross-LLM robustness. We evaluate our method on two alternative LLM backends, DeepSeek-V3,2 and Qwen-Plus, to test its robustness across models with distinct generative styles and capabilities. We focus on ECS for its controlled and reproducible setting across LLMs. As shown in Figure 8(a), coordination-based coalescing preserves the NCI, polarization, and GD trajectories of both Qwen-Plus and DeepSeek, closely following their dynamical opinion trends. This demonstrates that our topology-driven mechanism is robust across diverse LLM backbones. Scalability. We gradually scaled the simulation to 5,000 agents (ECS-5000), far beyond normal operating sizes in prior LLM-agent work. As shown in Figure 8(b), coordination still preserves global dynamics on a large scale. And token usage is substantially reduced across all tested scales, with up to 77.35% savings as reported in Table 3, demonstrating stable scalability in scaled networks.
5
6
CONCLUSION
In this work, we propose TopoSim, a topology-aware framework for LLM-driven social simulation that treats network structure as an active modeling signal rather than a passive scaffold. By explicitly modeling how topology shapes update coordination and influence, our approach enables scalable and realistic simulations. Experiments show that TopoSim achieves behavioral fidelity comparable to full agent simulations while reducing token usage by 50–90%, with strong generalization on real-world data. These results suggest a principled pathway for integrating graph-theoretic structure with LLM-based multi-agent systems. Future work may extend topology-aware dynamics to evolving networks, addressing the current limitation on static structures.
RELATED WORK
Social Simulation. Social simulation studies how social phenomena emerge from individual interactions within a social network. A 8
Topology-Aware LLM-Driven Social Simulation: A Unified Framework for Efficient and Realistic Agent Dynamics
REFERENCES [1] Sinan Aral and Dylan Walker. 2012. Identifying influential and susceptible members of social networks. Science 337, 6092 (2012), 337–341. [2] Eytan Bakshy, Jake M. Hofman, Winter A. Mason, and Duncan J. Watts. 2011. Everyone’s an influencer: quantifying influence on twitter. In Proceedings of the Fourth ACM International Conference on Web Search and Data Mining (Hong Kong, China) (WSDM ’11). Association for Computing Machinery, New York, NY, USA, 65–74. https://doi.org/10.1145/1935826.1935845 [3] Garrett Bernstein and Kyle O’Brien. 2013. Stochastic agent-based simulations of social networks. In Proceedings of the 46th Annual Simulation Symposium (San Diego, California) (ANSS 13). Society for Computer Simulation International, San Diego, CA, USA, Article 5, 8 pages. [4] Anwoy Chatterjee, H S V N S Kowndinya Renduchintala, Sumit Bhatia, and Tanmoy Chakraborty. 2024. POSIX: A Prompt Sensitivity Index For Large Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2024, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 14550–14565. https://doi.org/10.18653/v1/2024.findings-emnlp.852 [5] Yun-Shiuan Chuang, Agam Goyal, Nikunj Harlalka, Siddharth Suresh, Robert Hawkins, Sijia Yang, Dhavan Shah, Junjie Hu, and Timothy Rogers. 2024. Simulating Opinion Dynamics with Networks of LLM-based Agents. In Findings of the Association for Computational Linguistics: NAACL 2024, Kevin Duh, Helena Gomez, and Steven Bethard (Eds.). Association for Computational Linguistics, Mexico City, Mexico, 3326–3346. https://doi.org/10.18653/v1/2024.findings-naacl.211 [6] Yun-Shiuan Chuang, Agam Goyal, Nikunj Harlalka, Siddharth Suresh, Robert Hawkins, Sijia Yang, Dhavan Shah, Junjie Hu, and Timothy Rogers. 2024. Simulating Opinion Dynamics with Networks of LLM-based Agents. In Findings of the Association for Computational Linguistics: NAACL 2024, Kevin Duh, Helena Gomez, and Steven Bethard (Eds.). Association for Computational Linguistics, Mexico City, Mexico, 3326–3346. https://doi.org/10.18653/v1/2024.findings-naacl.211 [7] Guillaume Deffuant, David Neau, Frédéric Amblard, and Gérard Weisbuch. 2000. Mixing beliefs among interacting agents. Adv. Complex Syst. 3 (2000), 87–98. https://api.semanticscholar.org/CorpusID:15604530 [8] Claire Donnat and Susan Holmes. 2018. Tracking network dynamics: A survey using graph distances. The Annals of Applied Statistics 12, 2 (2018), 971–1012. [9] Joshua M. Epstein and Robert L. Axtell. 1996. Growing Artificial Societies: Social Science from the Bottom Up. The MIT Press. https://doi.org/10.7551/mitpress/ 3374.001.0001 [10] Stephen Eubank, Hasan Guclu, V. S. Anil Kumar, Madhav V. Marathe, Aravind Srinivasan, Zoltán Toroczkai, and Nan Wang. 2004. Modelling disease outbreaks in realistic urban social networks. Nature 429, 6988 (01 May 2004), 180–184. https://doi.org/10.1038/nature02541 [11] Neil M. Ferguson, Derek A.T. Cummings, Simon Cauchemez, Christophe Fraser, Steven Riley, Aronrag Meeyai, Sopon Iamsirithaworn, and Donald S. Burke. 2005. Strategies for containing an emerging influenza pandemic in Southeast Asia. Nature 437, 7056 (01 Sep 2005), 209–214. https://doi.org/10.1038/nature04017 [12] Antonino Ferraro, Antonio Galli, Valerio La Gatta, Marco Postiglione, Gian Marco Orlando, Diego Russo, Giuseppe Riccio, Antonio Romano, and Vincenzo Moscato. 2025. Agent-Based Modelling Meets Generative AI in Social Network Simulations. In Social Networks Analysis and Mining, Luca Maria Aiello, Tanmoy Chakraborty, and Sabrina Gaito (Eds.). Springer Nature Switzerland, Cham, 155–170. [13] Noah Friedkin and Eugene Johnsen. 1990. Social Influence and Opinions. Journal of Mathematical Sociology - J MATH SOCIOL 15 (01 1990), 193–206. https: //doi.org/10.1080/0022250X.1990.9990069 [14] Chen Gao, Xiaochong Lan, Zhihong Lu, Jinzhu Mao, Jinghua Piao, Huandong Wang, Depeng Jin, and Yong Li. 2023. S3: Social-network simulation system with large language model-empowered agents. arXiv preprint arXiv:2307.14984 (2023). [15] Chen Gao, Fengli Xu, Xu Chen, Xiang Wang, Xiangnan He, and Yong Li. 2024. Simulating Human Society with Large Language Model Agents: City, Social Media, and Economic System. In Companion Proceedings of the ACM Web Conference 2024 (Singapore, Singapore) (WWW ’24). Association for Computing Machinery, New York, NY, USA, 1290–1293. https://doi.org/10.1145/3589335.3641253 [16] Rainer Hegselmann and Ulrich Krause. 2002. Opinion Dynamics and Bounded Confidence Models, Analysis and Simulation. Journal of Artificial Societies and Social Simulation 5 (07 2002). [17] Keith Henderson, Brian Gallagher, Tina Eliassi-Rad, Hanghang Tong, Sugato Basu, Leman Akoglu, Danai Koutra, Christos Faloutsos, and Lei Li. 2012. RolX: structural role extraction & mining in large graphs. In Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (Beijing, China) (KDD ’12). Association for Computing Machinery, New York, NY, USA, 1231–1239. https://doi.org/10.1145/2339530.2339723 [18] Guiyang Hou, Wenqi Zhang, Zhe Zheng, Yongliang Shen, and Weiming Lu. 2025. Scaling LLMs’ Social Reasoning: Sprinkle Cognitive “Aha Moment” into Fundamental Long-thought Logical Capabilities. In Findings of the Association for Computational Linguistics: ACL 2025, Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (Eds.). Association for Computational Linguistics, Vienna, Austria, 3126–3138. https://doi.org/10.18653/v1/2025.
findings-acl.162 [19] Glen Jeh and Jennifer Widom. 2003. Scaling personalized web search. In Proceedings of the 12th international conference on World Wide Web. 271–279. [20] Zexun Jiang, Yafang Shi, Maoxu Li, Hongjiang Xiao, Yunxiao Qin, Qinglan Wei, Ye Wang, and Yuan Zhang. 2024. Casevo: A Cognitive Agents and Social Evolution Simulator. arXiv:2412.19498 [cs.SI] https://arxiv.org/abs/2412.19498 [21] Unchitta Kan, Michelle Feng, and Mason A Porter. 2023. An adaptive boundedconfidence model of opinion dynamics on networks. Journal of Complex Networks 11, 1 (2023), 415–444. https://doi.org/10.1093/comnet/cnac055 [22] Yuhan Liu, Zirui Song, Juntian Zhang, Xiaoqing Zhang, Xiuying Chen, and Rui Yan. 2025. The Stepwise Deception: Simulating the Evolution from True News to Fake News with LLM Agents. arXiv:2410.19064 [cs.SI] https://arxiv.org/abs/ 2410.19064 [23] Francois Lorrain and Harrison C White. 1971. Structural equivalence of individuals in social networks. The Journal of mathematical sociology 1, 1 (1971), 49–80. [24] FRANÇOIS LORRAIN and HARRISON C. WHITE. 1977. STRUCTURAL EQUIVALENCE OF INDIVIDUALS IN SOCIAL NETWORKS. In Social Networks, Samuel Leinhardt (Ed.). Academic Press, 67–98. https://doi.org/10.1016/B978-0-12442450-0.50012-2 [25] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. 1999. The PageRank citation ranking: Bringing order to the web. Technical Report. Stanford infolab. [26] 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 (San Francisco, CA, USA) (UIST ’23). Association for Computing Machinery, New York, NY, USA, Article 2, 22 pages. https://doi.org/10.1145/3586183.3606763 [27] Jinghua Piao, Yuwei Yan, Jun Zhang, Nian Li, Junbo Yan, Xiaochong Lan, Zhihong Lu, Zhiheng Zheng, Jing Yi Wang, Di Zhou, et al. 2025. AgentSociety: LargeScale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society. arXiv preprint arXiv:2502.08691 (2025). [28] Leonardo F.R. Ribeiro, Pedro H.P. Saverese, and Daniel R. Figueiredo. 2017. struc2vec: Learning Node Representations from Structural Identity. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’17). ACM, 385–394. https://doi.org/10.1145/3097983.3098061 [29] Chenxi Wang, Zongfang Liu, Dequan Yang, and Xiuying Chen. 2025. Decoding Echo Chambers: LLM-Powered Simulations Revealing Polarization in Social Networks. In Proceedings of the 31st International Conference on Computational Linguistics, Owen Rambow, Leo Wanner, Marianna Apidianaki, Hend Al-Khalifa, Barbara Di Eugenio, and Steven Schockaert (Eds.). Association for Computational Linguistics, Abu Dhabi, UAE, 3913–3923. https://aclanthology.org/2025.colingmain.264/ [30] Duncan J. Watts and Peter Sheridan Dodds. 2007. Influentials, Networks, and Public Opinion Formation. Journal of Consumer Research 34, 4 (05 2007), 441– 458. https://doi.org/10.1086/518527 arXiv:https://academic.oup.com/jcr/articlepdf/34/4/441/17929486/34-4-441.pdf [31] Steven H. Watts, Duncan J.and Strogatz. 1998. Collective dynamics of ‘smallworld’ networks. Nature 393, 6684 (01 Jun 1998), 440–442. https://doi.org/10. 1038/30918 [32] Harrison C White, Scott A Boorman, and Ronald L Breiger. 1976. Social structure from multiple networks. I. Blockmodels of roles and positions. American journal of sociology 81, 4 (1976), 730–780. [33] Ziyi Yang, Zaibin Zhang, Zirui Zheng, Yuxian Jiang, Ziyue Gan, Zhiyu Wang, Zijian Ling, Jinsong Chen, Martz Ma, Bowen Dong, Prateek Gupta, Shuyue Hu, Zhenfei Yin, Guohao Li, Xu Jia, Lijun Wang, Bernard Ghanem, Huchuan Lu, Chaochao Lu, Wanli Ouyang, Yu Qiao, Philip Torr, and Jing Shao. 2025. OASIS: Open Agent Social Interaction Simulations with One Million Agents. arXiv:2411.11581 [cs.CL] https://arxiv.org/abs/2411.11581 [34] Junchi Yao, Hongjie Zhang, Jie Ou, Dingyi Zuo, Zheng Yang, and Zhicheng Dong. 2025. Social opinions prediction utilizes fusing dynamics equation with LLM-based agents. Scientific Reports 15, 1 (03 May 2025), 15472. https://doi.org/ 10.1038/s41598-025-99704-3 [35] Haofei Yu, Zhaochen Hong, Zirui Cheng, Kunlun Zhu, Keyang Xuan, Jinwei Yao, Tao Feng, and Jiaxuan You. 2025. ResearchTown: Simulator of Human Research Community. arXiv:2412.17767 [cs.CL] https://arxiv.org/abs/2412.17767 [36] Albert Ĺaszló Barabási and Réka Albert. 1999. Emergence of scaling in random networks. Science 286 5439 (1999), 509–12. https://api.semanticscholar.org/ CorpusID:260558609