EpiiSLM: Clarifying Epidemic Modeling Methods

• EpiiSLM is an ambiguous label that conflates EpiILM’s mechanistic Bayesian inference with Epi-LLM’s behavior-driven epidemic simulation.
• EpiILM employs discrete-time individual-level models with spatial and network kernels, leveraging R and Fortran for efficient Bayesian inference.
• Epi-LLM integrates large language models with agent-based SEIR simulations to capture dynamic quarantine decisions and behavioral adaptations.

Searching arXiv for the exact term and close variants to disambiguate "EpiiSLM". "EpiiSLM" does not correspond, in the supplied arXiv materials, to a single standardized method name. The nearest attested names are EpiILM, an R package for Bayesian inference and simulation in discrete-time individual-level infectious-disease transmission models (V. et al., 2020); Epi-LLM, a framework that embeds LLMs in epidemiological agent-based simulations (Ferenz et al., 1 Jun 2026); and SLM (“Star Log-extended eMulation”), a reduced-order emulator for the Tolman-Oppenheimer-Volkoff equations in neutron-star modeling (Lalit et al., 2024). The most plausible interpretation, on lexical grounds, is that “EpiiSLM” is a misspelling or conflation of one or more of these names. In the epidemiological literature represented here, the closest exact referent is EpiILM, while Epi-LLM is the closest recent framework combining epidemic simulation with adaptive agent behavior. This suggests that the term is best treated as an ambiguous label rather than a canonical technical designation.

1. Disambiguation and scope

Within the supplied record, no paper is titled “EpiiSLM.” The corpus instead contains several near matches with distinct domains and methodological commitments. EpiILM denotes an R package for “simulation from, and inference for, discrete-time individual-level models of infectious disease transmission” in a Bayesian framework via Metropolis-Hastings MCMC (V. et al., 2020). Epi-LLM denotes a framework for probing LLM behavioral priors through epidemiological agent-based models (Ferenz et al., 1 Jun 2026). SLM denotes a logarithmic Dynamic Mode Decomposition-based emulator for neutron-star structure calculations (Lalit et al., 2024).

A concise comparison is useful because the ambiguous label can otherwise invite category errors.

Name	Domain	Core function
EpiILM	Infectious-disease statistics	Simulation and Bayesian inference for individual-level epidemic models
Epi-LLM	Epidemiological agent-based modeling	LLM-driven behavioral adaptation in outbreak simulations
SLM	Computational astrophysics	Efficient emulation of TOV solutions and tidal deformability

The infectious-disease interpretation is the most natural when “Epi” is taken literally. A plausible implication is that “EpiiSLM” arose from confusion between EpiILM and another acronym ending in SLM or LLM. The remainder of this entry therefore centers on the two epidemiological frameworks, while noting the distinct and unrelated astrophysical meaning of SLM.

2. EpiILM: individual-level epidemic modeling

EpiILM is an R package implementing the discrete-time individual-level modeling framework of Deardon et al. (2010), with inference in a Bayesian framework via Metropolis-Hastings MCMC and performance-critical routines coded in Fortran (V. et al., 2020). Its design target is data in which the disease status of each individual is tracked over time and infection risk depends on the current state of other individuals. This dependence distinguishes the framework from standard survival or multi-state models, because one person’s infection time is not independent of others’ infection times once the infectious pool evolves.

The package supports both SI and SIR compartmental settings. For a susceptible individual $i$ at time $t$, the discrete-time infection probability is

$$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$

where $\Omega_S(i)$ is the susceptibility function, $\Omega_T(j)$ is the transmissibility function, $\kappa(i,j)$ is the infection kernel, $I(t)$ is the set of infectious individuals at time $t$, and $\varepsilon$ is a spark term for infection from outside the observed population or from unmodeled sources (V. et al., 2020). The framework uses linear covariate effects for susceptibility and transmissibility. Susceptibility may include an intercept, whereas transmissibility is typically specified without one to avoid identifiability issues.

The likelihood is constructed from the full epidemic history. At each time step, it multiplies the probability of each new infection by the probability that all individuals remaining susceptible avoid infection. This full-history construction is central to the package’s mechanistic character: it encodes transmission as a process over interacting individuals rather than as marginal event times (V. et al., 2020).

3. Transmission kernels, inference, and software structure in EpiILM

EpiILM supports two principal classes of infection kernel. In spatial models, transmission decays with Euclidean distance according to

$\kappa(ij)=d_{ij}^{-\beta},$

with $t$0 the Euclidean distance and $t$1 governing spatial decay (V. et al., 2020). In network-based models, transmission follows one or more contact matrices,

$t$2

where $t$3 is the $t$4 entry of the $t$5-th weighted adjacency matrix. The package admits directed or undirected networks, weighted or unweighted, with the diagonal set to zero (V. et al., 2020).

Bayesian inference is performed with the epimcmc() function. The posterior has the standard form

$t$6

with independent marginal priors selected from the Gamma, Half-normal, or Uniform families (V. et al., 2020). The proposal mechanism is a Gaussian random walk. Proposal variances can be set manually, and fixed parameters can be imposed by setting their proposal variance to zero. The package also supports adaptive MCMC via adaptMCMC; if adapt = TRUE, a target acceptance rate can be supplied through acc.rate, but initial values and proposal variances remain required (V. et al., 2020).

The software is not restricted to estimation. Its main user-facing functions include epiBR0, epidata, plot.epidata, epidic, epilike, epimcmc, summary.mcmc, plot.mcmc, pred.epi, and plot.pred.epi (V. et al., 2020). This makes EpiILM simultaneously a simulation, visualization, likelihood-evaluation, and posterior-prediction environment.

A notable implementation detail is that expensive likelihood evaluation and core numerical routines are coded in Fortran and called from R. This matters because individual-level epidemic likelihoods can be computationally intensive inside MCMC, especially for large populations or long epidemic histories (V. et al., 2020).

4. Worked examples and intended use of EpiILM

The paper demonstrates EpiILM on both simulated and real datasets. A spatial SI example on 100 individuals uses

$t$7

with baseline susceptibility $t$8, spatial decay $t$9, and no spark term (V. et al., 2020). A network SI example on a directed binary network with a susceptibility covariate $$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$0 uses

$$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$1

The paper also presents a spatial SIR farm model in which susceptibility depends on the number of animals $$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$2,

$$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$3

and a network SIR example with both susceptibility and transmissibility covariates,

$$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$4

In the latter case, one susceptibility parameter is fixed to handle non-identifiability (V. et al., 2020).

The real-data case study concerns a greenhouse outbreak of tomato spotted wilt virus on 520 pepper plants. The analysis uses reconstructed spatial coordinates and infection/removal times, fits an SIR spatial model, and reports posterior estimates for $$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$5 and $$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$6. According to the paper, the example required about 10 minutes of computation on a standard laptop (V. et al., 2020).

The package’s intended niche is explicit in the paper. It is presented as filling a gap because few R packages support mechanistic individual-level epidemic modeling with both simulation and Bayesian inference, while simultaneously allowing spatial and network transmission, covariates on susceptibility and transmissibility, and posterior prediction (V. et al., 2020). The stated limitations are equally clear: it is a work in progress, currently focused on SI and SIR models, with possible future extensions including time-varying networks, time-varying covariates or spatial kernels, joint spatial-plus-network kernels, uncertain transition times, unknown covariates, additional compartmental frameworks such as SEIR and SIRS, and multiple strains or pathogens (V. et al., 2020).

5. Epi-LLM: LLM-driven behavioral adaptation in epidemic ABMs

If “EpiiSLM” was intended to refer not to a classical inference package but to a newer behavioral epidemic simulator, the nearest candidate is Epi-LLM (Ferenz et al., 1 Jun 2026). Epi-LLM integrates agent-based modeling, real-life epigames, and LLMs into a synthetic society in which agents reason and adapt dynamically over an outbreak contact network. Its stated motivation is that epidemics are shaped not only by pathogen biology but also by human behavior, especially voluntary actions such as quarantine and distancing, while classical compartmental models like SEIR are too coarse to represent heterogeneous individuals, dynamic contact networks, adaptive decision-making, and behavioral feedback loops (Ferenz et al., 1 Jun 2026).

The framework is built on Starsim, an open-source agent-based infectious-disease platform. It couples a dynamic contact network, an SEIR disease process, and LLM-driven daily quarantine decisions (Ferenz et al., 1 Jun 2026). At initialization, agents receive health-belief attributes from survey data, the contact network is generated from a fitted log-normal degree distribution, and an initial disease prevalence of 1% is seeded. Each day, an agent observes health status, local symptomatic prevalence among contacts, current point balance, and a prompt containing belief scores and context, then produces a binary decision: quarantine or stay active (Ferenz et al., 1 Jun 2026).

The simulation duration is 15 days, matching the AUIB epigame. The disease process is a modified SEIR with asymptomatic, mild, and severe states. The paper reports epidemiological parameters including infectiousness $$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$7 per hour, exposed-to-infected progression $$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$8 per hour, recovery rate $$\P(i,t) =1- \exp\{-\Omega_S(i) \sum_{j \in I(t)}{\Omega_T(j) \kappa(ij)}-\varepsilon\}, \qquad \Omega_S(i) >0,\ \Omega_T(j)>0,\ \varepsilon > 0$$9 per hour, and death rates of $\Omega_S(i)$0, $\Omega_S(i)$1, and $\Omega_S(i)$2 for asymptomatic, mild, and severe cases with proportions $\Omega_S(i)$3, $\Omega_S(i)$4, and $\Omega_S(i)$5, respectively (Ferenz et al., 1 Jun 2026). The contact network uses Starsim’s RandomNet module with median contact duration 10 seconds and dynamic rewiring.

The framework is validated against the AUIB epigame study, with Health Belief Model dimensions mapped to infection risk, health severity, quarantine response efficacy, and self-efficacy plus group assignment as an economic-cost proxy (Ferenz et al., 1 Jun 2026). The tested LLM architectures are DeepSeek V3, Llama 3 70B, Nemotron 120B, and GPT-OSS 120B. A self-agreement score is defined as

$\Omega_S(i)$6

with $\Omega_S(i)$7 indicating full determinism and $\Omega_S(i)$8 indicating a random split between yes and no responses (Ferenz et al., 1 Jun 2026).

6. Quantitative findings and conceptual significance of Epi-LLM

The principal quantitative result is that LLM agents across four architectures reduced peak active infections relative to a no-intervention SEIR baseline, with quarantine compliance peaking at 58–65% on day six of the 15-day simulation (Ferenz et al., 1 Jun 2026). In geographically framed GPT-OSS conditions, the no-intervention trajectory peaked at about 300 active infections, whereas the geographically labeled conditions peaked later and lower, around 200 active infections (Ferenz et al., 1 Jun 2026).

The geographic-framing experiment assigns agents one of four labels—China, Iraq, Kenya, or United Kingdom—to test whether labels alone induce culturally differentiated behavior. The reported result is negative: final cumulative infections ranged from 436 (Iraq; attack rate 95.6%) to 447 (UK; attack rate 98.0%), a difference of only 11 agents. The paper further reports a Kruskal–Wallis test on daily quarantine rates of $\Omega_S(i)$9, non-significant pairwise Mann–Whitney tests after Bonferroni correction, and Pearson correlations between cumulative infection trajectories of $\Omega_T(j)$0 (Ferenz et al., 1 Jun 2026). The authors conclude that geographic labels alone do not produce meaningful differentiation; explicit attitudinal parameterization is required.

Behavioral analysis is conducted with GLMs. A binomial GLM for the proportion of days an agent quarantined achieved Cox–Snell pseudo-$\Omega_T(j)$1, compared with 0.072 in the human trial (Ferenz et al., 1 Jun 2026). The strongest predictor was perceived health severity, with

$\Omega_T(j)$2

The reported additive-model coefficients are: Intercept $\Omega_T(j)$3, $\Omega_T(j)$4; Group (B vs A) $\Omega_T(j)$5, $\Omega_T(j)$6; Perceived infection risk $\Omega_T(j)$7, $\Omega_T(j)$8; Perceived health severity $\Omega_T(j)$9, $\kappa(i,j)$0; Quarantine self-efficacy $\kappa(i,j)$1, $\kappa(i,j)$2; and Quarantine response efficacy $\kappa(i,j)$3, $\kappa(i,j)$4 (Ferenz et al., 1 Jun 2026). A Poisson GLM with survival offset finds Group B significantly lower in quarantine rate ($\kappa(i,j)$5), infection risk strongly positive ($\kappa(i,j)$6), health severity strongly positive ($\kappa(i,j)$7), self-efficacy negative ($\kappa(i,j)$8), response efficacy non-significant ($\kappa(i,j)$9), and no significant interactions with group (Ferenz et al., 1 Jun 2026).

Architectural variation is treated as substantively important rather than incidental. Llama 3 70B is described as highly self-consistent and near-perfect in self-agreement, while Nemotron 120B exhibits lower self-agreement around $I(t)$0 and more stochastic behavior (Ferenz et al., 1 Jun 2026). The paper interprets low-variance architectures as offering greater internal validity for testing behavioral rules and high-variance architectures as possibly better approximating real-world decision-making. This suggests that, within Epi-LLM, architecture acts as a parameter of social simulation rather than merely an implementation choice.

7. Relation to SLM and the status of “EpiiSLM” as a term

The remaining nearby acronym in the supplied materials is SLM, short for Star Log-extended eMulation, a method for efficient computation of the Tolman-Oppenheimer-Volkoff equations in neutron-star structure and tidal-deformability calculations (Lalit et al., 2024). SLM is methodologically unrelated to epidemic modeling. It applies Dynamic Mode Decomposition in a logarithmic representation of pressure, mass, and tidal variables to obtain an emulator that yields a speed-up of approximately $I(t)$1 relative to high-fidelity Runge–Kutta integration, with maximum-mass errors below $I(t)$2 and radii within $I(t)$3 on the reported tabular equations of state (Lalit et al., 2024). Its appearance alongside EpiILM and Epi-LLM is best understood as lexical rather than conceptual proximity.

Accordingly, “EpiiSLM” is best classified as an ambiguous, non-canonical label. In the infectious-disease context, the strongest exact referent is EpiILM (V. et al., 2020). In the recent literature on adaptive epidemic simulation with language-model agents, the closest conceptual referent is Epi-LLM (Ferenz et al., 1 Jun 2026). The most defensible editorial treatment is therefore disambiguation rather than forced normalization to a single method name.

This suggests two distinct research lineages behind the apparent typo. One lineage is mechanistic statistical epidemiology, represented by EpiILM: discrete-time transmission kernels, full epidemic-history likelihoods, Bayesian inference, and Fortran-accelerated computation in R (V. et al., 2020). The other is behaviorally enriched epidemic simulation, represented by Epi-LLM: agent-based SEIR dynamics, survey-grounded belief parameterization, and LLM-mediated quarantine decisions (Ferenz et al., 1 Jun 2026). Both address epidemic processes at the individual level, but they differ sharply in epistemic target. EpiILM is designed for inference on transmission mechanisms from epidemic data, whereas Epi-LLM is designed to probe behavioral priors and intervention-sensitive dynamics in synthetic societies.