DynamiX: Dynamic Social Network Simulator

• DynamiX is a large-scale dynamic social network simulator that models adaptive role switching and endogenous network rewiring.
• It employs a modular design with dynamic hierarchy selection, LLM-driven core agents, and inequality-oriented behavior for ordinary agents.
• Experimental evaluations demonstrate improved attitude evolution alignment, enhanced link prediction, and novel insights on polarization and follower growth.

DynamiX is a large-scale social network simulator introduced for dynamic social network modeling, with the explicit aim of moving beyond static-graph simulations and modeling two coupled processes that characterize real social platforms: which users act as “core” opinion drivers at a given moment, and how follow–unfollow relations evolve under multi-dimensional social forces (Sun et al., 26 Jul 2025). The system combines a dynamic hierarchy module, an information-stream-based link-prediction method for opinion leaders, and an inequality-oriented behavior and link-prediction engine for ordinary agents. In the reported experiments, it improves attitude evolution alignment and collective behavior analysis relative to static-network baselines, while also supporting follower-growth prediction and robustness studies for dynamic link prediction (Sun et al., 26 Jul 2025).

1. Conceptual scope and research setting

DynamiX is motivated by the claim that existing large-language-model-based social simulations have mainly emphasized scaling agent populations while neglecting the dynamic evolution of social relationships (Sun et al., 26 Jul 2025). Its central premise is that fidelity in social simulation depends not only on modeling agent cognition or message generation, but also on representing adaptive role switching and endogenous rewiring of the follow graph. In this formulation, social influence is not assigned to a fixed elite set, nor treated as exogenous; it is recalculated at each timestep.

The simulator is therefore built around two key dynamics. First, it determines which users should be treated as “core” agents at time $t$, reflecting adaptive switching of user roles. Second, it models how individual links are created or removed in response to stance similarity, trust, exposure, influence, and asymmetry in social interactions (Sun et al., 26 Jul 2025). This suggests a shift from static diffusion over a frozen topology to co-evolutionary simulation, in which opinions and network structure recursively affect one another.

Within the broader literature on evolving social networks, DynamiX addresses a different problem from formal influence-optimization work. For example, “Dynamic Influence Maximization” studies how to maintain a seed set $S_t$ of size at most $k$ in an evolving network so as to maximize expected influence spread, proving a $(1-1/e-\epsilon)$-approximation in the incremental model and SETH-based hardness in the fully dynamic model (Peng, 2021). DynamiX, by contrast, is a simulator of social interaction and network rewiring rather than an algorithm for maintaining an optimal seed set.

2. System architecture and representational design

The architecture of DynamiX comprises an Environment Server together with three pluggable modules (Sun et al., 26 Jul 2025). The Environment Server stores user personas, memories, timelines, and the current follow graph. This stateful substrate is then used by the Dynamic Hierarchy selector, the Core-Agent decision unit, and the Inequality-Oriented Behavior & Link-Prediction engine for ordinary agents.

This modular decomposition is technically important because it separates persistent social state from role selection and action generation. Personas and memories support continuity of agent behavior; timelines encode local exposure; and the current follow graph serves as the substrate on which link prediction and rewiring operate. The design also allows the computationally expensive LLM-based simulation to be concentrated on the subset of agents currently designated as core.

For core agents, the decision prompt integrates persona $\mathcal{P}_i$, memory $\mathcal{M}_P^i$, environment $\mathcal{M}_E^i$, and a personalized external information stream $R_i$ (Sun et al., 26 Jul 2025). For ordinary agents, the simulator uses an augmented agent-based model rather than fully LLM-mediated deliberation. This architectural asymmetry reflects the model’s assumption that high-impact users and ordinary users should not be simulated with the same decision mechanism.

3. Dynamic hierarchy and the modeling of opinion leaders

At each timestep $t$, DynamiX computes for every agent $i$ an influence–diversity score $S_t$0 and promotes the top-$S_t$1 agents to core status (Sun et al., 26 Jul 2025). The score combines spread potential and content diversity:

$S_t$2

$S_t$3

Core membership is written as a role-adaptation function,

$S_t$4

or, more generally,

$S_t$5

This dynamic selection contrasts both fixed-set and random baselines, ensuring that agents with the greatest reach and diversity drive the LLM-based core simulation at each round (Sun et al., 26 Jul 2025). The combination of reach and diversity is notable: high influence alone is not sufficient, because diversity of exposure is also used to determine which agents matter for the next step of simulation.

For these core agents, DynamiX introduces an information-stream-based link-prediction method. Beyond their follower timeline, they receive a personalized “outside” information stream $S_t$6 consisting of candidate tweets from friends-of-friends and high-reach users (Sun et al., 26 Jul 2025). Each candidate $S_t$7 is scored as

$S_t$8

where $S_t$9 is agent $k$0’s last-tweet embedding, $k$1 is the embedding of candidate tweet $k$2, $k$3 is the decay rate of timeliness, and $k$4 captures aggregate influence such as likes, retweets, and follower count (Sun et al., 26 Jul 2025). In abstraction, the score can be written as

$k$5

where $k$6 is embedding-based similarity and $k$7 summarizes recency and influence. Core agents then choose top-$k$8 candidates above a similarity threshold and decide whether to follow or unfollow.

A plausible implication is that DynamiX treats opinion-leader behavior as jointly driven by attitudinal homophily, timeliness, and exogenous exposure to high-salience content, rather than by local neighborhood structure alone.

4. Ordinary agents, inequality, and adaptive rewiring

Ordinary agents are modeled through an inequality-oriented behavior decision-making module and a multi-factor link-prediction engine (Sun et al., 26 Jul 2025). The framework assumes that ordinary users do not indiscriminately absorb influence from all neighbors. Instead, only sufficiently trusted and like-minded neighbors contribute to attitude updates:

$k$9

Attitudes are then updated by

$(1-1/e-\epsilon)$0

and the heterogeneous interaction weight is

$(1-1/e-\epsilon)$1

Here $(1-1/e-\epsilon)$2 is a trust score propagated via two-hop paths:

$(1-1/e-\epsilon)$3

The model also gives an abstract relationship-adjustment form,

$(1-1/e-\epsilon)$4

although the concrete implementation is the $(1-1/e-\epsilon)$5-weighted trust-plus-influence formulation above (Sun et al., 26 Jul 2025).

Link rewiring for ordinary agents is performed by separate missing-link and spurious-link scores. The missing-link score is

$(1-1/e-\epsilon)$6

while the spurious-link score is

$(1-1/e-\epsilon)$7

With probabilities $(1-1/e-\epsilon)$8 and $(1-1/e-\epsilon)$9, ordinary agents periodically choose the $\mathcal{P}_i$0 maximizing $\mathcal{P}_i$1 or $\mathcal{P}_i$2 to follow or unfollow (Sun et al., 26 Jul 2025). This mechanism is explicitly intended to capture asymmetric, attitudinally driven rewiring and unequal social interactions.

The main conceptual point is that DynamiX does not represent network change as random perturbation. Rewiring is conditioned on stance proximity, reverse connectivity, trust, and exposure asymmetry, thereby embedding social inequality directly into the simulator’s update rules.

5. Experimental evaluation and reported findings

DynamiX is evaluated on three real-world Twitter event collections—Moon Landing Conspiracy with 3.3 K users and 6.7 K tweets, Xinjiang Cotton with 9.9 K users and 14.2 K tweets, and Trump–Russia Investigation with 7.9 K users and 10.3 K tweets—as well as the public Congress network for link-prediction benchmarking (Sun et al., 26 Jul 2025). The reported metrics are attitude evolution alignment metrics $\mathcal{P}_i$3Bias, $\mathcal{P}_i$4Div, Dynamic Time Warping, and Fréchet distance, together with link-prediction Precision, Recall, and F1.

On macro-alignment, DynamiX cuts the next-best $\mathcal{P}_i$5Bias in half, with the example $\mathcal{P}_i$6 versus $\mathcal{P}_i$7, and achieves approximately $\mathcal{P}_i$8 lower DTW and Fréchet distance consistently across all three events (Sun et al., 26 Jul 2025). The reported interpretation is that dynamic rewiring improves agreement with real attitude trajectories relative to static-network simulation.

The dynamic-network evaluation uses 100 K agents with 2 K core agents. In this setting, polarization intensifies sharply under dynamic rewiring: by $\mathcal{P}_i$9, the fraction of extreme attitudes rises from $\mathcal{M}_P^i$0 to $\mathcal{M}_P^i$1, a $\mathcal{M}_P^i$2 rise, compared with $\mathcal{M}_P^i$3 in a static graph (Sun et al., 26 Jul 2025). New follows cluster homogeneously by stance, with neutral agents bridging clusters, and a micro-case study shows how the sequential influence of new ties, personalized streams, and timeline posts can drive an agent from neutral to extreme support of euthanasia.

DynamiX is also used for follower-growth prediction under three interventions—tweet frequency, content quality, and trending promotion. The reported findings are that high-influence users gain most from trending promotion, low-influence users require sustained high-quality content, and frequency alone has limited effect for either group (Sun et al., 26 Jul 2025). The paper frames this as opening a new theoretical perspective on follower growth prediction and providing empirical evidence for opinion leaders cultivation.

For link-prediction robustness, across edge-perturbation rates from $\mathcal{M}_P^i$4 to $\mathcal{M}_P^i$5, DynamiX outperforms common heuristics including CN, AA, Katz, and LPOD by $\mathcal{M}_P^i$6 in F1 for both missing-link and spurious-link tasks (Sun et al., 26 Jul 2025). Ablation and sensitivity analyses further show that removing the dynamic hierarchy or the inequality-weighted ABM raises $\mathcal{M}_P^i$7Bias by up to $\mathcal{M}_P^i$8, replacing the DLPE with classical predictors drops link precision by $\mathcal{M}_P^i$9, and performance degrades beyond the reported optimum $\mathcal{M}_E^i$0 and $\mathcal{M}_E^i$1.

6. Broader significance, applications, and disambiguation

DynamiX is presented as advancing the fidelity of large-scale social simulators by explicitly weaving together dynamic role switching, attitudinal homophily, trust inequality, and content timeliness (Sun et al., 26 Jul 2025). The paper identifies several application areas: studying collective polarization under dynamic rewiring, theoretically grounding opinion-leader cultivation, scenario testing for misinformation, norm-change, or advertising campaigns, and policy design around echo-chamber mitigation. Future directions include incorporating multi-modal signals such as images and video, richer behavioral repertoires such as direct messages and group chats, and fine-grained style control over LLM agents (Sun et al., 26 Jul 2025).

A recurring source of confusion is nomenclature. The name “DynamiX” or near variants has also been used for unrelated systems, including “DynaMix” for zero-shot dynamical-systems reconstruction (Hemmer et al., 19 May 2025), “DynaMIX” for mixed-precision resource optimization of real-time DNN applications (Cho et al., 2023), “DYNAMIX” for PPO-based adaptive batch-size optimization in distributed machine learning (Dai et al., 9 Oct 2025), “DynamiX” for dynamic resource exploration in ad recommendation (Roychowdhury et al., 23 Nov 2025), and “$\mathcal{M}_E^i$2YNAMICS” for language-based rigid-body dynamics inference from video (Kao et al., 20 May 2026). In the social-simulation literature, however, DynamiX specifically denotes the large-scale simulator for dynamic social network modeling introduced in 2025 (Sun et al., 26 Jul 2025).

Taken as a whole, DynamiX formalizes a co-evolutionary view of online social systems: influential users are not fixed, ordinary users are not socially symmetric, and network topology is not a passive background variable. That framing is its main theoretical contribution, and the reported experiments suggest that explicitly modeling those dependencies changes both the quality of attitude-evolution alignment and the qualitative dynamics of polarization, follower growth, and social clustering (Sun et al., 26 Jul 2025).