PaperlayerView exact revision on arXiv

Source-linked AI summary

LLMs for Social Network Modeling: From Network Generation to Dynamic Processes

Shikha Mallick, Alex Thomo, Akrati Saxena

arXiv:2609.08049v1cs.SIcs.AI

TL;DR

LLM-based social network modeling lacks a unified perspective across studies, despite its use for representing network formation and social dynamics. This survey organizes the literature by modeling mechanism and compares approaches, concluding that LLM agents support richer contextual social modeling while current methods face realism, scalability, bias, reproducibility, and validation challenges.

  • Problem

    Related surveys do not systematically organize how LLM agents generate social network structure or model evolving opinions, information, and rumors through networked interaction.

    Source: §I · pp. 1–2

  • Method

    The survey reviews LLM-based social network modeling by network generation and dynamic processes, comparing representative methods, evaluation, and reproducibility information.

    Source: §I · p. 2

  • Results

    Across the reviewed areas, LLM agents enable richer representations of users, communication, and context-dependent social behavior than classical rule-based models.

    Source: §V · p. 7

  • Takeaways & Limitations

    The review identifies behavioral realism, scalability, bias, reproducibility, and validation against real social systems as limitations of current methods.

    Source: §V · p. 7

    Current methods face behavioral-realism, scalability, bias, reproducibility, and real-world validation limitations.

    Source: §V · p. 7

Abstract

from arXiv · show

Large language models (LLMs) are rapidly emerging as a new paradigm for modeling social networks by representing users and their relationships and interactions through natural language. Unlike classical network models or deep learning approaches, LLMs can simulate context-aware social behavior and language-driven interactions, enabling more realistic modeling of network formation and dynamic social processes. However, existing studies are scattered across different research communities and lack a unified perspective. This survey presents the first comprehensive review of LLMs for social network modeling by organizing the literature into two broad categories: network generative models and dynamic process models. Network generative models are further classified into selection-based and interaction-based approaches, while dynamic process models are categorized into opinion dynamics, information diffusion, and rumor propagation, each with their underlying modeling mechanisms. LLMs enable rich textual social interactions and decision-making, but they also exhibit many limitations, including inherent social biases and prompt sensitivity. We outline these open research challenges and discuss future directions in LLM-based social network modeling.

I. INTRODUCTION

Classical and deep-learning approaches offer efficient or data-driven models but remain limited in representing context-dependent communication and adaptive behavior. This survey provides a unified review of LLM-based social network modeling, organizing work around network generation and dynamic processes and identifying open challenges.

I. INTRODUCTION

  • Classical models rely on predefined rules, while deep-learning approaches require task-specific data and latent representations and remain limited in open-ended, context-dependent social behavior.Classical models are efficient, interpretable, and theoretically grounded; the survey presents both approaches as having limitations for modeling adaptive social behavior.Source: §I · p. 1
  • LLM agents can form networks and simulate social dynamics through messages, opinion updates, behavior adoption, and information or rumor propagation.Their use also raises challenges in realism, scalability, bias, reproducibility, and validation against real-world behavior.Source: §I · p. 1
  • The survey organizes LLM-based social network modeling into network generative models and dynamic process models, including opinion dynamics, information diffusion, and rumor propagation.It compares representative studies by LLM backbone, dataset, fine-tuning strategy, evaluation protocol, and code availability.Source: §I · p. 2
  • Prior surveys cover adjacent areas but do not systematically organize how LLM agents generate network structure or model evolving opinions, information, and rumors.The survey frames its contribution as a unified, mechanism-centered review of these modeling approaches.Source: §I · p. 2

II. LLMS FOR NETWORK GENERATIVE MODELS

LLM network generators either select edges directly or let them emerge through repeated interactions. In selection-based models, prompts and conditioning choices shape generated topology, with studies reporting both structural patterns and social biases.

A. Selection-based Network Generative Models

  • Selection-based models use a fixed node set, with an LLM selecting edge targets based on network structure, node attributes, or both.The survey distinguishes structure-conditioned from attribute-conditioned approaches.Source: A. Selection-based Network Generative… · p. 2
  • Repeated sequential choices produced heavy-tailed, approximately scale-free degree distributions, but name preferences also generated highly centralized networks.New nodes received existing node names and degrees before selecting a fixed number of connection targets.Source: A. Selection-based Network Generative… · p. 3
  • Complete neighborhood descriptions produced more realistic heavy-tailed structures than degree-only prompts, which often produced centralized or star-like networks.The experiments also found strong preferences for mutual neighbors and socially similar nodes.Source: A. Selection-based Network Generative… · p. 3
  • GraphMind infers missing edges from node profiles and serialized multi-hop paths, improving structural fidelity while potentially inheriting training-network biases.The resulting synthetic networks can serve as benchmarks for graph-learning and bot-detection algorithms.Source: A. Selection-based Network Generative… · p. 3
  • LLM edge choices favored preferential attachment but substantially overestimated political and religious homophily when attribute similarity and degree informed selection.The experiments varied one categorical social attribute at a time while always including node degree as a structural feature.Source: A. Selection-based Network Generative… · p. 3
  • Local and sequential prompting improved connectivity, clustering, and degree heterogeneity over global prompting, while sequential prompting better captured long-tailed degree patterns.All three prompting strategies overestimated political homophily, and repeatedly including all personas limited scalability.Source: A. Selection-based Network Generative… · p. 3
  • Varying revealed attributes in sequential generation produced strong demographic homophily, personality-related popularity effects, and other potential social biases.Candidate targets came from two-hop neighbors and randomly sampled non-neighbors.Source: A. Selection-based Network Generative… · p. 3
  • Across 192 networks generated from 50 personas, design choices substantially affected outcomes; smaller models made qualitatively different friendship choices, and political homophily was overestimated.The networks showed realistic clustering and modularity across the tested cultures, languages, model scales, and prompting methods.Source: A. Selection-based Network Generative… · p. 3

B. Interaction-based Network Generative Models

Interaction-based generators form edges iteratively through repeated node interactions. The reviewed approaches use evolving user-content activity or persistent agents’ relationship updates, and the survey recommends evaluating both resulting topology and edge-formation validity.

B. Interaction-based Network Generative Models

  • Interaction-based models derive edges iteratively from repeated node interactions over a fixed number of rounds.The nodes are LLM agents that interact and form connections across rounds.Source: B. Interaction-based Network… · pp. 3–4
  • GraphAgent-Generator builds evolving actor-item bipartite networks from user actions on content, which can be projected into citation, co-authorship, recommendation, and follow networks.Its framework does not systematically test whether prompting produces the heterogeneous agent behaviors needed for realistic network formation.Source: B. Interaction-based Network… · p. 4
  • Persistent agents’ communications, evaluations, and reflections can update relationship weights so edges emerge or weaken over time.This approach requires many LLM calls and calibration of rewards, update rules, and edge thresholds.Source: B. Interaction-based Network… · p. 4
  • Evaluation should assess both final topology and edge-formation validity, including sensitivity to prompts, model families, and fine-tuning strategies.The surveyed generators differ in direct edge selection versus edges emerging from prior behavior.Source: B. Interaction-based Network… · p. 4

III. LLMS AS DYNAMIC PROCESS MODELS

Dynamic process models track how LLM-agent states and actions change through networked interaction. Opinion-dynamics studies examine how conversational context and social influence shape individual and collective opinions.

III. LLMS AS DYNAMIC PROCESS MODELS

  • Dynamic process models simulate how LLM-agent states and actions change over time through networked interaction.The survey groups them into opinion dynamics, information diffusion, and rumor propagation.Source: §III · p. 4

A. Opinion Dynamics Models

  • Opinion-dynamics models use initial stances, memory, personas, and conversational context to simulate belief changes that can produce consensus, fragmentation, or polarization.These models capture persuasive content and heterogeneous responses, but may reproduce pretraining and alignment biases rather than realistic social behavior.Source: A. Opinion Dynamics Models · p. 4

1) Dialogue-based Opinion Dynamics Models:

  • Dialogue-based models update opinions through conversational reasoning, memory, or visible discussion histories.Without induced cognitive bias, agents reached fact-aligned consensus; prompted confirmation bias produced persistent disagreement, while other studies found selective persuasion and self-reinforcing majority dominance.Source: 1) Dialogue-based Opinion Dynamics… · p. 4

2) Exposure-based Opinion Dynamics Models:

Opinion-dynamics models examine how exposure, formal theory, and co-evolution shape LLM agents’ beliefs and collective outcomes. Across these approaches, network structure, model bias, and reciprocal feedback can affect trajectories and endpoints.

2) Exposure-based Opinion Dynamics Models:

  • Exposure to a wider range of viewpoints more often produced convergence, but effects also depended on persona design and initial opinion distributions.The reviewed work also links selective attention to herding and finds that defenses reduce, but do not eliminate, polarization from adversarial agents.Source: 2) Exposure-based Opinion Dynamics… · p. 5

3) Theory-based Opinion Dynamics Models:

  • Theory-based models compare or interpret LLM opinion trajectories using formal dynamics or empirically calibrated behavioral models.Reported outcomes include consensus rates related to network structure, final positions differing from DeGroot predictions, and variation in self-trust and susceptibility across LLM families.Source: 3) Theory-based Opinion Dynamics Models · p. 5

4) Co-evolution-based Opinion Dynamics Models:

  • Co-evolution models let opinions evolve jointly with communication structure or semantic environments, producing reciprocal feedback between beliefs and subsequent interactions or topics.SPARK showed stance shifts affect emerging subtopics, which subsequently reshape agent stances.Source: 4) Co-evolution-based Opinion Dynamics… · p. 5
  • Opinion updating can represent the content and framing of social influence, but relies on an opaque, model-dependent mechanism.The survey recommends evaluating trajectories and population outcomes alongside robustness tests and validation against human behavior.Source: 4) Co-evolution-based Opinion Dynamics… · p. 5

B. Information Diffusion Models

LLM-based information diffusion models simulate how content or behaviors spread through networks, while rumor models address belief and changes in transmitted messages. The survey distinguishes these processes by their modeling mechanisms.

B. Information Diffusion Models

  • Information diffusion begins with seed nodes that adopt or share information, exposing neighbors who may then activate and continue the spread.LLM-based models add language-aware actors that interpret content and generate context-dependent reactions.Source: B. Information Diffusion Models · p. 5

1) Content-based Information Diffusion Models:

  • Content-based models simulate posts moving through networks, using individual agents, content generators, or aggregate groups.Approaches include LLM-controlled posts paired with formal engagement or proxy diffusion models, and linguistic reactions combined with classical influence diffusion; group agents reduce costs but may hide within-group differences.Source: 1) Content-based Information Diffusion… · pp. 5–6

2) Adoption-based Information Diffusion Models:

  • Adoption-based models study agents adopting behaviors, policies, technologies, cooperation, or norms through social reinforcement.Studies report lower adoption thresholds than human participants and show unfair policy interventions can suppress cooperation and propagate norm erosion.Source: 2) Adoption-based Information… · p. 6

C. Rumor Propagation Models

1) Belief-based Rumor Propagation Models:

Rumor-propagation research examines belief, message transformation, and interventions. The surveyed work spans susceptibility and correction, factual degradation, malicious amplification, and fact-checking, with evaluation extending beyond whether rumors spread.

1) Belief-based Rumor Propagation Models:

  • Belief-based models simulate whether agents accept, reject, discuss, or retransmit existing rumors or false claims.FPS used persona-conditioned agents with memory and reflection; findings include greater effectiveness of early repeated correction and dependence of cascade size and intervention effectiveness on network structure.Source: 1) Belief-based Rumor Propagation Models · p. 6

2) Transformation-based Rumor Propagation Models:

  • Transformation-based models track semantic distortion, claim loss, and factual degradation as messages pass between users.FUSE varies spreader, commentator, and verifier roles, while a question-answering auditor tracks claims surviving persona-conditioned rewrites.Source: 2) Transformation-based Rumor… · p. 6

3) Intervention-based Rumor Propagation Models:

  • Intervention-based models examine malicious or corrective actions such as poisoning, censorship, reporting, and fact-checking.TrendSim studies attacks on trending topics; MOSAIC compares community, third-party, and hybrid fact-checking, reporting lower rumor susceptibility for safety-aligned LLMs than human users.Source: 3) Intervention-based Rumor… · p. 6

IV. OPEN CHALLENGES AND FUTURE DIRECTIONS

  • Evaluation of rumor models should measure reach, belief changes, factual loss, moderation success, and human realism.The survey identifies open challenges and future directions for making LLM-based social network models robust.Source: 3) Intervention-based Rumor… · p. 6, §IV · p. 6

A. Validation, Benchmarking, and Reproducibility

The survey identifies validation and reproducibility gaps alongside challenges in scalability, fairness, and temporal integration for LLM-based social-network simulations.

A. Validation, Benchmarking, and Reproducibility

  • Plausible individual decisions can still yield unrealistic collective patterns, and behavioral similarity does not establish causal validity; results also vary with models, prompts, and configurations.The passage calls for benchmarks combining real network structures with timestamped interactions and reports of prompts, update rules, computational costs, and simulation settings.Source: A. Validation, Benchmarking, and… · pp. 6–7

B. Scalable Simulation

  • Repeated LLM calls across agents and rounds make large simulations costly; group-agent and selective-reasoning approaches reach million-agent experiments but aggregation may hide heterogeneity.The passage identifies multi-fidelity simulation as a direction for balancing computational cost with population representativeness.Source: B. Scalable Simulation · p. 7

C. Fairness, Bias, and Responsible Simulation

  • Biased agent choices can distort connectivity, exposure, and influence; future work should address fairness, privacy, and persuasion risks while preserving realistic behavior and structure.The survey points to fairness-aware models for network formation, opinion dynamics, and influence analysis.Source: C. Fairness, Bias, and Responsible… · p. 7

D. Dynamic and Co-evolving Networks

  • Most systems capture only part of the feedback among ties, user states, content, and platform exposure; long-run reliability and temporal coverage remain limited.The passage calls for a unified framework for tie formation and dissolution, evolving identities and memories, and platform-mediated exposure.Source: D. Dynamic and Co-evolving Networks · p. 7

E. Theory-Guided Foundation Models

The survey presents theory-guided LLM approaches as a way to combine testable social-dynamics mechanisms with contextual interaction, while noting limits in scope and open questions for hybrid human-LLM networks.

E. Theory-Guided Foundation Models

  • Hybrid approaches combine formal social-dynamics models with LLMs for testable mechanisms and contextual semantics, but remain task-specific and concentrated on opinion dynamics.The passage contrasts these approaches with general-purpose models whose opaque, variable update rules limit interpretability, reproducibility, and transfer.Source: E. Theory-Guided Foundation Models · p. 7
  • Future social-network foundation models should use multimodal interaction data and constraints from network science, behavioral evidence, and causal social theories.The passage describes these as directions for future work.Source: E. Theory-Guided Foundation Models · p. 7

F. Hybrid Human-LLM Social Networks

  • Early studies suggest hybrid human-LLM networks may show collective behaviors distinct from purely human or AI systems, while existing models may not capture their heterogeneous interactions.Future work should examine network formation, asymmetric influence, evolving social roles, and co-evolution of network structure and behavior.Source: F. Hybrid Human-LLM Social Networks · p. 7

V. CONCLUSION

The survey organizes LLM-based social-network modeling around network generation and dynamic processes, synthesizing its capabilities, limitations, and future research directions.

V. CONCLUSION

  • The survey organizes prior work into network generative models and dynamic process models, covering selection and interaction approaches, opinion dynamics, information diffusion, and rumor propagation.It presents this as a unified, mechanism-centered review.Source: §V · p. 7
  • LLM agents support richer representations of users, communication, and context-dependent social behavior than classical rule-based models.Source: §V · p. 7
  • Current methods face limitations in behavioral realism, scalability, bias, reproducibility, and validation against real social systems.Source: §V · p. 7
  • The survey highlights future directions in benchmarking, scalable simulation, co-evolving networks, fairness-aware modeling, theory-guided approaches, and hybrid human-LLM networks.It concludes that progress will likely depend on network science, behavioral validation, and responsible simulation practices.Source: §V · p. 7

SUPPLEMENTARY MATERIAL

The supplement provides reproducibility information for reviewed studies and explains how to read its model, fine-tuning, dataset, and code-availability entries.

SUPPLEMENTARY MATERIAL

  • Tables S1 and S2 report reviewed studies’ LLM backbones, fine-tuning choices, real datasets, and public-code availability.The tables are divided into parts, including Table S1 Part I and Table S2 Part II.Source: SUPPLEMENTARY MATERIAL · p. 11, Table S1 · p. 11, Table S2 · p. 12
  • Boldface marks either a study’s sole evaluated task LLM or its author-identified best-performing LLM; no boldface means no unique aggregate winner or rankings dependent on metric or condition.Source: SUPPLEMENTARY MATERIAL · p. 11
  • FT means fine-tuning, ∗ marks parameter-efficient fine-tuning, ✓ and ✗ indicate whether fine-tuning was used, Optional allows both configurations, and – marks no applicable real dataset or unreported public code.Source: SUPPLEMENTARY MATERIAL · p. 11

Related papers

Loading 2609.08049v1…