Large language models for spreading dynamics in complex systems

Abstract: Spreading dynamics is a central topic in the physics of complex systems and network science, providing a unified framework for understanding how information, behaviors, and diseases propagate through interactions among system units. In many propagation contexts, spreading processes are influenced by multiple interacting factors, such as information expression patterns, cultural contexts, living environments, cognitive preferences, and public policies, which are difficult to incorporate directly into classical modeling frameworks. Recently, LLMs have exhibited strong capabilities in natural language understanding, reasoning, and generation, enabling explicit perception of semantic content and contextual cues in spreading processes, thereby supporting the analysis of the different influencing factors. Beyond serving as external analytical tools, LLMs can also act as interactive agents embedded in propagation systems, potentially influencing spreading pathways and feedback structures. Consequently, the roles and impacts of LLMs on spreading dynamics have become an active and rapidly growing research area across multiple research disciplines. This review provides a comprehensive overview of recent advances in applying LLMs to the study of spreading dynamics across two representative domains: digital epidemics, such as misinformation and rumors, and biological epidemics, including infectious disease outbreaks. We first examine the foundations of epidemic modeling from a complex-systems perspective and discuss how LLM-based approaches relate to traditional frameworks. We then systematically review recent studies from three key perspectives, which are epidemic modeling, epidemic detection and surveillance, and epidemic prediction and management, to clarify how LLMs enhance these areas. Finally, open challenges and potential research directions are discussed.

• The paper develops a taxonomy of LLM applications across digital and biological epidemics, covering modeling, surveillance, prediction, and intervention while treating LLM agents as both analytical tools and active spreaders.
• The review reports gains including PandemicLLM’s 119.7% forecasting improvement over the CDC ensemble, RumorSphere’s 64% reduction in opinion deviation, and outbreak correlations up to 0.82, but highlights high computational costs and sparse validation.
• The paper identifies practical priorities for trustworthy deployment: standardized benchmarks, empirical calibration, improved sensitivity for early warning, bias and hallucination controls, and theory linking language-mediated interactions to epidemic thresholds and stability.

This paper, published as a review in Physics Reports, synthesizes the emerging literature on how LLMs reshape the study of spreading dynamics in complex systems (2602.08085). The authors—spanning network science, statistical physics, and public health institutions—organize a rapidly growing body of work around two canonical propagation domains, digital epidemics (rumors, misinformation, ideological diffusion) and biological epidemics (influenza, dengue, Ebola, SARS-CoV-2), and three task families: epidemic modeling, detection and surveillance, and prediction and management. Their central thesis is twofold: LLMs serve both as analytical instruments that make semantic, cultural, cognitive, and policy factors explicitly modelable, and as endogenous participants in spreading processes, introducing new feedback loops that classical compartmental and agent-based frameworks cannot represent.

Motivation and positioning relative to classical frameworks

The review opens from the standard statistical-physics formulation of spreading on networks: nodes in discrete states ($S$, $I$, $R$), transitions governed by rates such as $\beta$ and $\gamma$, with macroscopic behavior controlled by the epidemic threshold $\lambda_c$. It contrasts this tradition—which spans SIS/SIR compartmental models, threshold and independent-cascade contagion models, and large-scale ABMs—with the observation that real-world propagation is modulated by information expression patterns, cultural context, cognition, and public policy, factors typically "sacrificed to preserve analytical tractability."

The authors argue LLMs change this situation along three axes. First, in representation, LLMs ingest heterogeneous, weakly structured signals (posts, clinical notes, reports) via prompting and retrieval without bespoke feature engineering. Second, in adaptation, they offer zero-shot/few-shot generalization and counterfactual role conditioning suited to non-stationary regimes. Third, in decision support, they produce human-readable rationales linking micro-level behaviors to macro-level dynamics. A second, more consequential claim is that LLM-driven agents are no longer merely observers: human–AI co-evolution couples model training data to subsequent human preferences, introducing nonlinearity into both information diffusion and disease-related behaviors.

Foundations and validation methodology

The background section provides primers on spreading dynamics (random, scale-free, small-world, temporal, multilayer, and higher-order networks; threshold vs. independent-cascade mechanisms) and on LLM technology (pre-training → instruction tuning → RLHF/RLAIF; causal-decoder, prefix-decoder, and encoder–decoder architectures; multimodal embedding of text, images, graphs, and time series). Notably, it also treats empirical validation systematically—a point of genuine methodological value. Empirical data play three roles: initialization (e.g., SNAP ego-networks, ANES personas), driving (e.g., OASIS's time engine activating agents on real behavioral schedules), and evaluation. Evaluation metrics span accuracy and macro-F1 for discrete states, DTW and Pearson correlation for trajectory alignment, and cosine similarity for semantic fidelity, applied at micro level (stance/content/behavior alignment per agent) and macro level (attitude distribution bias and diversity).

A candid weakness is acknowledged here: current studies use empirical validation only sparingly, and no standardized benchmark suite yet exists. The authors compile candidate datasets (Twitter15/16, PHEME, FakeNewsNet, Weibo21, MuMiN, Weibo-COV, etc.) to partially address this gap.

LLM-based epidemic modeling

Digital epidemics

The modeling literature divides into LLM agents replacing rule-based ABM agents and LLMs acting as exogenous drivers of spreading.

On the first front, early work by Li et al. used simple persona prompts over random, scale-free, and high-brokerage networks, finding fake news spreads fastest in scale-free topologies and slowest where brokerage bottlenecks exist, with extraversion and openness predicting sharing propensity. FPS extended this with short-term memory, long-term memory, and reflection modules, reporting that political misinformation propagates significantly faster than misinformation about terrorism, disasters, science, folklore, or finance, and that agreeableness and neuroticism correlate with susceptibility. Chuang et al. introduced confirmation bias into speaker–listener opinion dynamics: without it, LLM agents converge rapidly to consensus; with it, polarization emerges, with stronger bias producing greater diversity. They also report an important caveat—LLM agents tend to reject inaccurate information and endorse factual content regardless of assigned roles, limiting their ability to simulate individuals who resist evidence.

Scalability is treated honestly. The review tabulates computational costs: GABM at 1K agents costs $20 and 80 h per round on CPU; AgentSociety runs 10K agents at ~459 s/round; OASIS reaches 1M agents but requires 24 A100 GPUs at 2.8 h/round. Against classical systems—EpiSimdemics simulating hundreds of millions of people in 12 seconds on Blue Waters—this is orders of magnitude slower. Hybrid architectures are the main mitigation: FDE-LLM assigns opinion leaders to LLM role-playing constrained by cellular automata while followers follow CA+SIR dynamics; RumorSphere dynamically reclassifies core vs. ordinary agents each step based on informational disorder, reducing average opinion deviation by 64% relative to existing models and implicating tightly connected local communities as key rumor accelerators. On network generation, GAG produces social graphs up to$10^5$nodes and$10^7$ edges that reproduce seven macroscopic properties including power-law degree distributions and shrinking diameters, entirely zero-shot.

On the second front—LLMs as endogenous spreaders—the review identifies three mechanisms: (i) LLMs as low-cost producers of near-human-quality content, functioning as continuously injected "pathogens" under inoculation theory, effectively raising the misinformation reproduction rate; (ii) LLM-controlled bots participating in human-centered diffusion, with evidence they exhibit lower adoption thresholds than humans and can act as super-spreaders; and (iii) spreading among populations composed purely of LLM agents, where De Marzo et al. observed emergent scale-free network formation and collective phenomena analogous to human societies—an instance of machine behavior in Rahwan's sense. Lu et al.'s randomized experiment with 176 participants is highlighted as particularly significant: LLM-driven adversarial comments reduced participants' true/false discrimination and suppressed sharing of authentic news, with effects moderated by political orientation. The authors interpret this as a directed, adaptive perturbation field rather than neutral noise—one that plausibly lowers the critical threshold for misinformation outbreaks—and argue future models must treat LLM bots as active propagators coupled to system stability and criticality, not as uniform stochastic disturbances.

Biological epidemics

Direct-involvement studies include Villaplana et al.'s ChatGPT-3.5 agents with heterogeneous attributes, which showed that richer epidemic information flattens case-growth curves and attenuates peaks—an explicit demonstration of behavioral feedback absent in SIR. EpiLLM autoregressively predicts spatiotemporal transmission segments with GNN-encoded epidemiological tokens; PandemicLLM and MIFlu fuse epidemiological time series, policy text, and genomic surveillance reports. Indirect involvement includes ChatGPT-assisted SEIR code generation (lowering the modeling barrier for non-programmers), LLM-enhanced spatially disaggregated simulation, and LLMob, which generates socially constrained mobility trajectories that reproduce realistic pandemic-period activity distributions by conditioning agents on government-response prompts.

Detection and surveillance

For digital perception, the review organizes methods into three targets:

1. Content and contexts: LESS4FD achieves over 17.5% accuracy and 16.1% F1 improvements over TextCNN/BERT-class baselines on MM-COVID via LLM-extracted topic–entity heterogeneous graphs; MAGE-fend gains over 6.8% accuracy/F1 over non-LLM video baselines on TikCron. Zero-shot self-detection findings are notable: LLMs detect machine-generated disinformation more reliably than human-written disinformation, and perform better on long news articles than short, noisy social posts.
2. Spreading networks: LLM-generated synthetic reactions reconstruct unobservable user–news interaction graphs (DELL); chain-of-clues prompting on pruned propagation networks addresses LLM context-length limits; LLM pseudo-labels stabilize global label propagation over cross-modal news networks.
3. Social cognition: a distinctive contribution of this review is framing stance evolution, emotion, and subjective network representations ("social sensing") as perceptual objects that LLMs can extract at scale, supplementing structure-only perception.

For biological surveillance, the review situates LLM-enhanced methods against established event-based systems (EIOS deployed in 160+ countries, EPIWATCH, BEACON, HealthMap, ProMED-mail). Quantitative results are strongest here: in conjunctivitis infoveillance across seven LLMs, GPT-4's outbreak probability correlated at 0.73 with expert labels, Mixtral-8×22B's outbreak-size estimates correlated at 0.82 with ground-truth case counts, and GPT-4 achieved specificity of 0.83–1.00 but only moderate sensitivity of 0.32–0.71 on validated infectious etiology—indicating LLMs recover multiple epidemiological dimensions rather than binary alarms, though false negatives remain a concern for early warning. On ongoing monitoring, Xie et al.'s fine-tuned pipeline over self-reported tweets recovered symptom trends consistent with variant emergence (e.g., taste/smell loss tracking the Beta variant), shortened recovery times 2020–2023, and quantified reinfection rates (fewer than 10% reinfected within 180 days; ~1% infected three times).

Prediction and management

Digital prediction methods fall into three classes: direct LLM forecasting (Wang et al.'s fine-tuned news-conditioned time-series framework with iterative reasoning-agent filtering); LLM-as-component approaches (CARE's cascade-retrieved in-context learning, AutoCas's autoregressive cascade tokenization, LoRA-tuned metadata-to-text transformation); and simulation-based forecasting via LLM agent societies (S³, Y Social, MOSAIC, LLM-AIDSim). Management studies evaluate intervention timing and frequency within simulated environments: FPS finds early fact-checking sharply reduces initial believers but decays through forgetting, mid-stage intervention yields sustained suppression at higher residual exposure, and daily versus every-three-days official corrections differ little—suggesting moderate-interval fact-checking is cost-effective. Li et al. compare encouraging commentary, disclosing inaccuracy post hoc, and blocking high-connectivity, high-openness influencers. The authors note candidly that LLM-based intervention research remains coarse relative to the traditional nudging/boosting/refutation taxonomy.

Biological forecasting shows the largest headline numbers in the review: PandemicLLM improves forecasting accuracy by 119.7% and reduces MSE by 72.4% relative to the CDC Ensemble Model by reformulating prediction as multimodal text reasoning over demographics, hospitalization series, policy documents, and variant reports. MIFlu achieves accurate two-week ILI forecasts while conceding that twenty-week horizons remain weak. These results carry direct policy implications: heterogeneous inputs inaccessible to mechanistic or statistical baselines materially improve short-horizon public-health decisions, though long-horizon reliability and interpretability constrain operational use.

Limitations and open questions

The review is explicit about constraints, several of which qualify the quantitative claims above. Computational cost limits LLM-agent simulations to thousands of agents per hour-scale rounds versus millions per second for optimized ABMs. Validation against empirical data remains sparse and non-standardized; results may be sensitive to specific model versions, prompts, and decoding strategies, threatening reproducibility. LLM agents display systematic behavioral biases (factual conformity, polarized judgments relative to human annotators, internalized ideological bias) that undermine their fidelity as proxies for resistant or biased populations. In biological surveillance, sensitivity as low as 0.32 implies missed outbreaks. Interpretability, demographic bias inherited from training corpora, adversarial vulnerability, and hallucination are flagged as unresolved risks directly affecting WHO-level intervention decisions. Open scientific questions include: principled mappings between language-based interaction patterns and analytically tractable dynamical variables; the stability, amplification, and controllability of spreading systems with endogenous LLM participants; incorporation of higher-order interactions beyond dyadic contact; and benchmarking frameworks that disentangle cognitive, content-driven, and structural contributions to observed outcomes.

Conclusion

This review consolidates a fragmented, fast-moving literature into a coherent taxonomy spanning digital and biological spreading, and makes a defensible conceptual argument: LLMs shift spreading dynamics research from content-blind state-transition models toward frameworks in which semantics, cognition, and generative AI agents are first-class dynamical components. Its strongest empirical anchors—PandemicLLM's forecast gains, RumorSphere's 64% deviation reduction, the 0.73–0.82 correlations in infoveillance—are offset by honest treatment of cost, validation deficits, and agent-fidelity problems. The most substantive open problem it identifies is theoretical rather than technical: how to couple language-mediated agent interactions back into the analytic machinery of statistical physics, so that LLM-era spreading can be characterized by thresholds, stability criteria, and phase behavior rather than by benchmark scores alone.