---
title: 'EpidemIQs: LLM-based Epidemic Modeling'
url: https://www.emergentmind.com/topics/epidemiqs
type: topic
---

# EpidemIQs: LLM-based Epidemic Modeling

Searching arXiv for the specified paper and closely related work to ground the encyclopedia entry.
EpidemIQs is a framework name used in multiple, distinct epidemic-intelligence contexts. In the 2025 paper "EpidemIQs: Prompt-to-Paper LLM Agents for Epidemic Modeling and Analysis" [2510.00024], it denotes a multi-agent LLM framework that integrates user inputs and autonomously conducts literature review, analytical derivation, network modeling, mechanistic modeling, stochastic simulations, data visualization and analysis, and finally documentation of findings in a structured manuscript. Earlier materials use the same name for a real-time epidemic-risk monitoring and early-detection system centered on the HiRES risk map and HiRES-p personal risk score [2004.00908], and for an integrated digital-epidemiology architecture spanning data ingestion, spatio-temporal analysis, infodemic monitoring, and policy optimization [2104.03611]. The 2025 usage is therefore best understood as a specific prompt-to-paper agent system within a broader lineage of computational epidemic-analysis frameworks [2510.00024].

## 1. Definition and architectural organization

The 2025 EpidemIQs framework comprises two agent types: scientist agents and task-expert agents [2510.00024]. Scientist agents are global orchestrators that decompose the user query into sub-tasks, plan the workflow, invoke tools and expert agents, reflect on interim outputs, and assemble the final report. Task-expert agents are specialized “tools” that perform atomic duties such as literature retrieval, network construction, mechanistic derivation, stochastic simulation, visualization, and LaTeX drafting [2510.00024].

Its interaction protocol is explicitly staged. The user sends an input query to the ScientistAgent; the ScientistAgent invokes PlanModule to produce a structured plan (JSON) of phases and sub-tasks; for each sub-task, the ScientistAgent issues a ReAct call, selecting a TaskExpert or internal tool, passing required inputs; the TaskExpert executes the duty and returns structured output; the ScientistAgent appends \( \text{Action}_t \), \( \text{Result}_t \) to short-term memory and invokes ReflectModule to analyze consistency/errors; if ReflectModule signals “revise,” the loop repeats; otherwise the workflow proceeds. On plan completion, the ScientistAgent collects artifacts and calls ReporterExpert and LatexCraftExpert to draft and validate the manuscript [2510.00024].

The supplied pseudo-code formalizes this as a bounded coordination loop with revision up to \(r_{\max}\), followed by a reporting phase in which `ReporterExpert.compose_report(artifacts)` and `LatexCraftExpert.validate_and_compile(Draft)` produce the final report [2510.00024]. This organization suggests that EpidemIQs is designed not merely as a conversational assistant but as a workflow engine with explicit planning, memory updates, reflective error checking, and manuscript compilation.

## 2. Modeling and simulation substrate

EpidemIQs incorporates mechanistic dynamical-system models, network models, and stochastic simulators as first-class components of the workflow [2510.00024]. The mechanistic layer includes the SIR system,
\[
dS/dt = -\beta SI/N,\qquad
dI/dt = \beta SI/N - \gamma I,\qquad
dR/dt = \gamma I,
\]
and the SEIR system,
\[
dS/dt = -\beta SI/N,\qquad
dE/dt = \beta SI/N - \sigma E,\qquad
dI/dt = \sigma E - \gamma I,\qquad
dR/dt = \gamma I.
\]
For a competitive-spread setting, the framework also includes an extended two-pathogen model with compartments \(S, I_1, I_2, R\) and interaction term \(\mu I_1 I_2\) [2510.00024].

The network-modeling layer is implemented via NetworkX and includes Erdős–Rényi random graphs \(G(N,p)\) with mean degree \(\langle k\rangle = p(N-1)\), Barabási–Albert scale-free networks with preferential attachment parameter \(m\), activity-driven temporal networks in which each node activates at rate \(\alpha\) and forms \(m\) edges per activation step, and custom multilayer networks for competitive spread [2510.00024]. Parameterization includes \(N\), average degree \(\langle k\rangle\), and power-law exponent for BA; vaccination scenarios can require tuning attachment rules to ensure sufficient nodes at target degree, for example degree \(=10\) [2510.00024].

The framework measures degree distribution, identification of hubs, and component sizes [2510.00024]. These modeling choices align with network-centric epidemic methods surveyed in digital epidemiology, where contact topology, mobility structure, and multi-scale interaction patterns are treated as primary determinants of transmission behavior [2104.03611]. A plausible implication is that EpidemIQs is meant to operationalize standard epidemic-theory primitives inside an LLM-directed execution loop rather than replace those primitives with purely text-native reasoning.

Its stochastic simulation layer centers on the Gillespie (direct) algorithm for continuous-time Markov chains. The state vector may take the form \(X=(S,E,I,R)\); reaction channels include infection, incubation, recovery, and vaccination; propensities include \(a_1=\beta SI\), \(a_2=\sigma E\), \(a_3=\gamma I\); the total rate is \(a_0=\sum_i a_i\); and the update draws \(\Delta t=-\ln(u)/a_0\), chooses reaction \(j\) with probability \(a_j/a_0\), then updates state and time until a stopping condition is met [2510.00024]. FastGEMF is used as an optimized engine for multilayer networks, and temporal networks use a custom discrete-time engine that updates adjacency each step [2510.00024].

## 3. Prompting, backbone models, and execution protocol

The 2025 implementation specifies distinct backbone LLMs by role. Scientist agents use OpenAI GPT-4.1 (2025-04-14), with a system prompt emphasizing planning, structured ReAct-Reflect loops, and JSON output schema. Task-expert agents use GPT-4.1-mini (2025-04-14), free-form except where schemas are enforced. MathExpert uses OpenAI o3-mini (2025-01-31), with chain-of-thought prompting for code generation and execution via Python sandbox to reduce hallucination [2510.00024].

Prompt engineering includes a predefined JSON schema for plans and reflections, tool declarations in system messages, and “Reminder” injection of the main query at each turn to avoid drift [2510.00024]. The protocol therefore combines explicit structure with iterative revision. This is technically notable because the framework’s reported control logic does not treat agent calls as unconstrained free-form dialogue; instead, plans and reflections are schema-constrained, while sub-task execution is tool-mediated [2510.00024].

The paper also defines a single-agent comparison condition: a single-agent LLM with the same system prompts and tools, iteratively planning, invoking tools, and revising outputs until task completion [2510.00024]. That design choice isolates the effect of multi-agent decomposition and specialist-role assignment from the effect of tool access alone. In the broader context of epidemiological QA, such decomposition is relevant because recent evidence indicates that epidemiological reasoning remains difficult for current LLMs, particularly on multi-step inference tasks [2601.03471]. This suggests that EpidemIQs’ emphasis on orchestration, reflection, and expert-role specialization is partly a response to the known brittleness of monolithic models on evidence-grounded epidemiological reasoning.

## 4. Experimental scenarios and quantitative performance

EpidemIQs is evaluated across five epidemic scenarios [2510.00024]. These are: effect of degree heterogeneity, using SEIR on ER vs BA with \(R_0=2.5\), \(\sigma=0.2\), \(\gamma=0.1\), seed \(=5\); reasons for transmission break, using SIR on ER/BA with \(R_0=3/0.5\), \(N=1000\), \(\langle k\rangle=10\); temporality effects, comparing activity-driven temporal and aggregated static networks with \(\alpha=0.1\), \(m=5\), \(R_0=3\); coexistence of two pathogens on a two-layer multiplex; and herd immunity via random vs targeted vaccination on an ER network engineered to have many degree-10 nodes [2510.00024].

The framework reports a completion success rate of \(100\%\), versus \(78 \pm 7.7\%\) for single-agent GPT-4.1 and \(80 \pm 6.3\%\) for single-agent o3 [2510.00024]. Human expert review by 5 blinded epidemiologists on a 1–10 rubric yields mean \(7.98 \pm 0.35\), and LLM-as-Judge yields mean \(9.04 \pm 0.21\) [2510.00024]. Review criteria are Readability & Writing Quality, Relevance & Motivation, Technical Soundness, Experimental Rigor, and Limitations & Ethics [2510.00024].

The average computation statistics are as follows:

| Phase | Time (s) | Tokens (K) |
|---|---:|---:|
| Discovery | 158 | 98 |
| Modeling | 140 | 175 |
| Simulation | 198 | 88 |
| Analysis | 177 | 93 |
| Report Writing | 517 | 416 |
| Overall | 1190 | 870 |

The same evaluation reports average cost per phase of \$0.10 for Discovery, \$0.72 for Modeling, \$0.23 for Simulation, \$0.16 for Analysis, \$0.36 for Report Writing, and \$1.57 overall, with cost computed via OpenAI May 2025 rate sheets applied to token usage per model type [2510.00024]. The framework’s end-to-end average token usage is 870 K tokens, and the autonomous process completed at a cost of about \$1.57 per study [2510.00024].

A direct comparison with the single-agent baselines shows the following:

| Model | Success % | Human Score |
|---|---:|---:|
| EpidemIQs | 100 | 7.98 |
| Single-GPT-4.1 | 78 ± 7.7 | 5.06 |
| Single-o3 | 80 ± 6.3 | 5.68 |

The same table reports time \(1190\) s, \(214\) s, and \(436\) s; tokens \(870\) K, \(312\) K, and \(165\) K; and cost \$1.57, \$0.91, and \$4.13 for EpidemIQs, Single-GPT-4.1, and Single-o3 respectively [2510.00024]. These figures indicate a trade-off: the multi-agent system uses more time and tokens than Single-GPT-4.1, but achieves higher success and higher human evaluation. A plausible implication is that the framework prioritizes completion reliability and report quality over minimal latency.

## 5. Reports, case studies, and relation to epidemic analytics

The framework consistently generated complete reports in scientific article format [2510.00024]. In Scenario 2, the paper cites a report snippet with Figure 1 showing time-series of \(S(t), I(t), R(t)\) for supercritical vs subcritical and for ER vs BA networks, and Table 3 reporting final epidemic sizes: ER with \(R_0=3\) gives \(885 \pm 25\), ER with \(R_0=0.5\) gives \(10 \pm 5\), BA with \(R_0=3\) gives \(520 \pm 40\), and BA with \(R_0=0.5\) gives \(8 \pm 6\) [2510.00024]. In Scenario 4, the workflow is summarized as analytical threshold derivation, multiplex network construction, Gillespie simulations, and regime classification plot for coexistence versus dominance [2510.00024].

These examples place EpidemIQs within a longer technical trajectory of epidemic-intelligence systems that combine modeling, simulation, and operational analytics. The HiRES/HiRES-p system, for example, defines an aggregated risk mean-field over communication-base-station cells and an individual risk score
\[
y_p(t_k)=\sum_{\ell,m} f_p(x^{\ell,m},t_k)\,\mathcal{R}(x^{\ell,m},t_k),
\qquad
\tilde{p}_p(t_k)=\max_{0\le i\le T_{\mathrm{inc}}} y_p(t_{k-i}),
\]
then applies either statistical inference or machine learning to flag suspected cases [2004.00908]. It reports that both detecting methods are above \(90\%\) accurate when the population infection rate is under \(20\%\), with the HiRES map covering \(\sim10^5\) cells and city-level risk correlating with official case counts at \(0.95\)–\(0.99\) [2004.00908]. By contrast, the 2025 EpidemIQs is not primarily a risk-scoring engine; it is a prompt-to-paper agent framework that can perform literature review, modeling, simulation, visualization, and article drafting [2510.00024].

Likewise, real-time epidemic monitoring methods based on daily cumulative counts \(C(t)\), \(R(t)\), and \(D(t)\) define active cases \(N_a(t)=C(t)-R(t)-D(t)\), mortality rate \(\mu(t)=\dot N_d(t)/N_a(t)\), recovery rate \(r(t)=\dot N_r(t)/N_a(t)\), and effective reproduction number
\[
R(t)=\frac{\dot N_T(t)}{\dot N_d(t)+\dot N_r(t)}
\]
for dashboard-style monitoring [2007.01508]. A plausible implication is that EpidemIQs could subsume such workflows as tasks or tool calls, but the 2025 paper does not claim direct incorporation of these particular formulas.

Forecasting systems furnish another point of comparison. Bayesian daily forecasting for regional COVID-19 outbreaks combines a two-timescale curve-fitting model and a 25-ODE SEIR-type model, re-estimated daily by adaptive Metropolis-Hastings, with anomaly detection based on predictive posterior breaches [2007.12523]. EWNet, by contrast, uses MODWT decomposition plus ARNN forecasting, with additive reconstruction
\[
\hat Y_{t+h}=\sum_{j=1}^J \hat D_{j,t+h}+\hat S_{J,t+h},
\]
and reports statistically significant long-term forecasting gains across 15 epidemic series [2206.10696]. EpidemIQs is different in aim: rather than proposing a single epidemic forecaster or monitor, it orchestrates multiple modeling and reporting steps into a complete manuscript pipeline [2510.00024].

## 6. Limitations, ambiguities, and future development

The 2025 paper identifies several limitations [2510.00024]. These include dependence on LLM capabilities and pricing; occasional minor hallucinations, such as mis-formatted tables and rare reference errors; limited depth in literature review at the abstract level; focus on network-based mechanistic models with no ABM or data-driven forecasting yet; and the continued necessity of human oversight to catch subtle errors, including an engine bug in temporal simulation [2510.00024]. These statements delimit the current scope of automation: the system is autonomous in execution flow, but not fully autonomous in epistemic validation.

Its listed future work is correspondingly concrete: extension to agent-based and individual-based epidemic models; integration of real-world mobility and case-report data for live outbreak forecasting; incorporation of full-text literature ingestion for deeper knowledge extraction; enhancement of agent ideation for generating novel research questions and hypotheses; and stronger ethical safeguards and provenance tracking in automated outputs [2510.00024]. These priorities are consistent with broader challenges in digital epidemiology, especially data bias, privacy, interoperability, and trust in human-centered epidemic technologies [2104.03611].

A terminological ambiguity surrounds the name itself. In the 2020 trajectory-data paper, “EpidemIQs” is described as a system for real-time, high-resolution epidemic-risk monitoring and early detection of suspected cases, consisting of HiRES and HiRES-p [2004.00908]. In the 2021 review-derived material, EpidemIQs is an integrated intelligence framework with DataIQ, ModelIQ, SpatioIQ, InfoIQ, PolicyIQ, and EthicsIQ modules [2104.03611]. In the 2025 paper, EpidemIQs is a prompt-to-paper multi-agent LLM system [2510.00024]. The shared name does not therefore denote a single stable architecture across the literature excerpted here. Instead, it denotes a family of epidemic-intelligence concepts whose common thread is the integration of computational methods for epidemic monitoring, modeling, or decision support. This suggests that the 2025 framework should be interpreted as the LLM-agent instantiation of a broader integrated-epidemic-intelligence idea, rather than as a direct continuation of one uniform prior system.

Source: https://www.emergentmind.com/topics/epidemiqs