EpiNarrate: Agentic Generation of Grounded Narratives from Epidemiological Scenario Projections

# EpiNarrate: Agentic Generation of Grounded Narratives from Epidemiological Scenario Projections
Source: [https://arxiv.org/html/2607.15544](https://arxiv.org/html/2607.15544)
Rituparna Datta Department of Computer Science University of Virginia Srini Venkatramanan Biocomplexity Institute University of Virginia Bryan L\. Lewis Biocomplexity Institute University of Virginia Yiqi Su Department of Computer Science Virginia Tech Harry Hochheiser Department of Biomedical Informatics University of Pittsburgh Lucie Contamin Department of Biomedical Informatics University of Pittsburgh Parantapa Bhattacharya Biocomplexity Institute University of Virginia Naren Ramakrishnan Department of Computer Science Virginia Tech Anil Vullikanti Biocomplexity Institute and Department of Computer Science University of Virginia

###### Abstract

Generation of clear and accessible public health narratives is critical for communicating complex epidemiological projections to policymakers and the general public at large\. Such narratives require more than simply reporting numbers: projections must be contextualized and quantitatively grounded across multiple dimensions\. Further, projections are often derived from large ensemble datasets which combine intervention assumptions, geographic and demographic strata, outcomes, time horizons, and uncertainty quantiles\. However, directly using large language models \(LLMs\) to summarize and contextualize such data often leads to inconsistencies, omissions, and fragile behavior\. We introduceEpiNarrate, an agentic framework for public health report generation that separates structured numerical reasoning from natural\-language generation\. The framework first extracts scenario axes and organizes them into a partial\-order schema, enabling systematic traversal of the underlying multidimensional space\. It then constructs an augmented dataset and derives valid quantitative statements through a comparison grammar that enforces semantic and arithmetic consistency\. To balance coverage and non\-redundancy, we introduce an interestingness\-driven selection mechanism based on maximum\-entropy principles\. Experiments on the COVID\-19 Scenario Modeling Hub demonstrate thatEpiNarrateproduces narratives with improved factual grounding and broader coverage of salient epidemiological patterns, while preserving the style of expert\-written reports\.

## 1Introduction

Public health agencies[CDC](https://arxiv.org/html/2607.15544#bib.bib9);[CDC](https://arxiv.org/html/2607.15544#bib.bib10);[CDC](https://arxiv.org/html/2607.15544#bib.bib11);[WHO](https://arxiv.org/html/2607.15544#bib.bib12)rely on carefully crafted, evidence\-backed, statements \(termed narratives\) to communicate key information to the broader public\. These can include information about ongoing and forecasted trends, e\.g\., Flu[CDC](https://arxiv.org/html/2607.15544#bib.bib11)or COVID\-19 incidence projections[CDC](https://arxiv.org/html/2607.15544#bib.bib10), and are typically written by domain experts to contextualize results from models or ensembles\. For instance, the COVID\-19 Scenario Modeling Hub \(SMH\) constructs multi\-model ensembles to provide robust, long\-term projections to guide federal and state\-level responsesMIDAS Network \([2025](https://arxiv.org/html/2607.15544#bib.bib21)\); Reichet al\.\([2022](https://arxiv.org/html/2607.15544#bib.bib31)\)\. Summary reports of these ensembles are produced[CDC](https://arxiv.org/html/2607.15544#bib.bib10), and are used by epidemiologists as well as the lay public\. Figure[1](https://arxiv.org/html/2607.15544#S1.F1)\(a\) shows an example of such a summary\.

Summary generation, as we study here, is a canonical task for LLMs and is closely related to the broader theme of data\-to\-text generation\. Prior work has studied this problem in settings such as table\-to\-text generation, where highlighted cells from Wikipedia tables are turned into faithful textual descriptions\(Parikhet al\.,[2020](https://arxiv.org/html/2607.15544#bib.bib23)\), and sports\-summary generation, where systems first select relevant records and plan their ordering before generating a coherent narrative\(Puduppullyet al\.,[2019a](https://arxiv.org/html/2607.15544#bib.bib25)\)\. Other work has compared modular pipeline systems with end\-to\-end neural generation, showing that explicit intermediate representations can improve generalization and output quality\(Ferreiraet al\.,[2019](https://arxiv.org/html/2607.15544#bib.bib24)\)\. However, public\-health scenario\-modeling reports differ from these standard settings because their inputs include large, multidimensional ensemble projections, scenario assumptions, geographic and demographic strata, and uncertainty quantiles\. Producing consumable reports requires not just summarizing the given data but placing them in context, grounding comparisons, and highlighting key takeaways\.

![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/inputOutput.png)\(a\)Example input\-output narrative
![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/data-flow.png)\(b\)Data flow Abstraction

Figure 1:\(a\) Example input and output fromEpiNarratefor an SMH round\. The input consists of structured projection data, while the output is a grounded narrative that reports scenario contrasts, burden, intervention effects, and age\-specific interpretation\. \(b\) Data flow inEpiNarratehas the following structure:Prompt↔\\;\\leftrightarrow\\;Partial Order→\\;\\rightarrow\\;Augmented dataset→\\;\\rightarrow\\;Set of facts→\\;\\rightarrow\\;Interesting facts\. The partial order is an abstract representation of the relationship between attribute values, and supports easier verification\. It is used to construct an augmented dataset and then a set of facts, using grammars\. Interesting facts are generated using an entropy based approach\. The components here map to those in the architecture ofEpiNarrate, shown in Figure[2](https://arxiv.org/html/2607.15544#S3.F2)The motivating question for our work is thus:*can agentic systems be developed to prepare narratives for public health information?*

We show that directly prompting even state\-of\-the\-art LLMs to generate summaries from epidemiological data leads to frequent structural inconsistencies and fragile behavior; this appears to be the case even under few shot learning\. We introduceEpiNarrate, an agentic framework for generating narratives for public health tasks; a crucial component inEpiNarrateis a partial order on some of the attributes, which helps improve the robustness of the generated summaries, as illustrated in Figure[2](https://arxiv.org/html/2607.15544#S3.F2)\. Our main contributions are:1\.We formalizeEpiNarrate, a public\-health narrative generation problem requiring correctness, coverage, and interestingness;2\.We proposeEpiNarrate, an agentic pipeline that uses partial\-order\-guided schema discovery, dataset augmentation, and comparison grammar to generate grounded facts, and3\.We introduce a MaxEnt based interestingness\-aware fact selection approach, which is evaluated by both agentic assessment and domain\-expert review\.

## 2Related Work

#### Data\-to\-text generation\.

Classical data\-to\-text \(D2T\) systems map structured records to text through hand\-designed pipelines for content selection, sentence planning, and surface realisationReiter and Dale \([2000](https://arxiv.org/html/2607.15544#bib.bib82)\)\. Neural D2TLebretet al\.\([2016](https://arxiv.org/html/2607.15544#bib.bib83)\); Wisemanet al\.\([2017](https://arxiv.org/html/2607.15544#bib.bib84)\); Parikhet al\.\([2020](https://arxiv.org/html/2607.15544#bib.bib23)\)replaced these stages with end\-to\-end sequence\-to\-sequence models trained on record\-and\-narrative pairs\. For example, WikiBioLebretet al\.\([2016](https://arxiv.org/html/2607.15544#bib.bib83)\)works on biographical infoboxes, RotoWireWisemanet al\.\([2017](https://arxiv.org/html/2607.15544#bib.bib84)\)on basketball box scores, and ToTToParikhet al\.\([2020](https://arxiv.org/html/2607.15544#bib.bib23)\)on highlighted Wikipedia table cells\. Castro Ferreira et al\.Ferreiraet al\.\([2019](https://arxiv.org/html/2607.15544#bib.bib24)\)compared end\-to\-end neural models against pipeline systems and reported that explicit intermediate representations produce more accurate output\. All of these methods were built for small flat tables that fit inside the encoder context window rather than the multi\-million\-cell hypercube that public\-health ensemble projections produce\.

Content selection and planning\.A separate strand of D2T argues that explicit modelling of*what to mention*, and*in what order*, improves faithfulness\. Puduppully et al\.Puduppullyet al\.\([2019a](https://arxiv.org/html/2607.15544#bib.bib25)\)introduced a neural pipeline for sports\-summary generation that selects records, orders them with a planner, and then realises them as prose\. Later work added hierarchical plannersRebuffelet al\.\([2020](https://arxiv.org/html/2607.15544#bib.bib85)\)and entity\-centric plansPuduppullyet al\.\([2019b](https://arxiv.org/html/2607.15544#bib.bib87)\)\. These planners were trained on small flat record tuples\. Adapting them to multi\-axis structured inputs requires substantial new engineering as done here\.

LLM prompting over structured inputs\.Recent work uses LLMs directly as generators, with the structured input flattened or summarised into the prompt\. Dibia’s LIDADibia \([2023](https://arxiv.org/html/2607.15544#bib.bib77)\)builds a column\-level statistical profile of a dataset such as dtypes, ranges, means, and sample values, and feeds that profile to the model as entire input\. Other work serialises tables into prompts as Markdown or CSV fragments and relies on long\-context models to reason over the linearised dataChen \([2023](https://arxiv.org/html/2607.15544#bib.bib90)\); Suiet al\.\([2024](https://arxiv.org/html/2607.15544#bib.bib96)\)\. A common limitation is that the model must decide which contrasts and quantities to surface with no structural constraint on validity, which can produce fabricated or mismatched comparisons\.

Structured summarization\.Some recent summarization work induces an intermediate structured representation between input and summary, including structured\-representation summarizationBalachandranet al\.\([2021](https://arxiv.org/html/2607.15544#bib.bib88)\)and graph\-aware summarizationKoncel\-Kedziorskiet al\.\([2019](https://arxiv.org/html/2607.15544#bib.bib89)\)\.EpiNarratealso uses an explicit intermediate structure\. It differs in two ways for the public\-health setting\. First, the intermediate representation is a poset over scenario axes induced from free\-text scenario descriptions, rather than a generic graph or knowledge structure\. Second, the fact\-generation step is governed by a domain\-aware comparison grammar that prevents structurally invalid contrasts\. We are not aware of any prior summarization system that imposes grammar of this kind, or that has been evaluated on hypercube\-scale ensemble projections\.

## 3Problem Statement

We consider the problem of automatically generating natural\-language summaries from large structured numerical datasets\. While this is a very general problem, we use public health as an exemplar application\. Let𝒟\\mathcal\{D\}denote a dataset organized alongKKdiscrete dimensionsℱ=\{a1,…,aK\}\\mathcal\{F\}=\\\{a\_\{1\},\\ldots,a\_\{K\}\\\}, where each dimensionaka\_\{k\}takes values in a finite set𝒱k\\mathcal\{V\}\_\{k\}\. Examples of such dimensions include intervention scenarios, outcome measures, population subgroups, geographic units, time periods, and uncertainty quantiles\. For example ifaka\_\{k\}denotes age group then𝒱k\\mathcal\{V\}\_\{k\}could be the groups\[0−10\],\[10−25\],⋯,\[65\+\]\[0\-10\],\[10\-25\],\\cdots,\[65\+\]\. Input𝒟\\mathcal\{D\}is often represented as a flat table, each record corresponds to a unique combination of dimension values\. Thus,𝒟\\mathcal\{D\}can be viewed as aKK\-way tensor,𝒟:𝒱1×⋯×𝒱K→ℝ,\\quad\\mathcal\{D\}:\\mathcal\{V\}\_\{1\}\\times\\cdots\\times\\mathcal\{V\}\_\{K\}\\rightarrow\\mathbb\{R\},where each entry𝒟​\[𝐯\]\\mathcal\{D\}\[\\mathbf\{v\}\]stores a numerical statistic associated with one configuration of the dimensions\.

![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/methodE2E.png)Figure 2:Pipeline for scenario\-based summary generation\. The framework converts projection data into structured scenario axes, builds a multidimensional hypercube, augments subset comparisons, filters valid facts using comparison grammar, MaxEnt and summarizes selected facts with LLM\.Problem statement\.Given𝒟\\mathcal\{D\}, and optionally a set of textual descriptions\{ms\}s∈𝒮\\\{m\_\{s\}\\\}\_\{s\\in\\mathcal\{S\}\}describing the scenarios, the goal is to generate a concise natural\-language summary𝒞​\(𝒟\)\\mathcal\{C\}\(\\mathcal\{D\}\)that satisfies three criteria: \(1\)Correctness: Every quantitative assertion in𝒞​\(𝒟\)\\mathcal\{C\}\(\\mathcal\{D\}\)should be grounded in a deterministic computation over𝒟\\mathcal\{D\}\. Formally, for each numerical claim, there should exist an executable computation pathπi\\pi\_\{i\}such that applyingπi\\pi\_\{i\}to𝒟\\mathcal\{D\}reproduces the stated value\. \(2\)Coverage\.The summary should cover the major dimensions of the dataset\. In addition, when reference summaries are available, the generated summary should recover the key information units expressed in the ground\-truth summaries associated with𝒟\\mathcal\{D\}, such as major scenario contrasts, high\-burden subgroups, regional heterogeneity, and peak or cumulative trends\. \(3\)Interestingness and non\-redundancy\.The objective is to choose a compact subset𝒞∗⊂𝒞\\mathcal\{C\}^\{\*\}\\subset\\mathcal\{C\}, with\|𝒞∗\|≪\|𝒞\|\|\\mathcal\{C\}^\{\*\}\|\\ll\|\\mathcal\{C\}\|, that is diverse, non\-redundant, and prioritizes comparisons with large effect sizes or surprising structure\. Redundant claims consume summary budget without improving the reader’s understanding of the dataset\.

For a given instance, we will output both𝒞\\mathcal\{C\}and𝒞∗\\mathcal\{C\}^\{\*\}\. The correctness requirement is critical because LLMs applied directly to large tables often hallucinate statistics, invert rankings, or report incorrect magnitudes\. The coverage requirement ensures that the summary is not limited to scenario\-level differences while omitting important variation across subgroups, regions, time, or uncertainty levels\.

Jointly satisfying all these is challenging\. Even a moderately sized setting with 5 scenarios, 3 outcomes, 4 subgroups, 50 regions, 52 weeks, and 23 uncertainty quantiles yields approximately1\.6×1061\.6\\times 10^\{6\}entries, far exceeding the context budget of current large language models\. Any summarization system must therefore perform selection, aggregation, or compression before generation, and errors at this intermediate stage directly constrain the quality of the final output\. Second, the combinatorial structure of𝒟\\mathcal\{D\}induces a large space of possible comparisons, making exhaustive enumeration and ranking impractical without a principled selection mechanism\. Consequently, naive end\-to\-end LLM approaches often fail along all three axes: they hallucinate or misstate numerical values, omit important dimensions, and overemphasize redundant or obvious findings\.

## 4Method

We propose a structured pipeline to generate grounded, quantitatively accurate summaries from scenario\-based ensemble projections\. The pipeline decomposes summary generation into five modular stages: \(*i*\) latent axis decomposition and dimension discovery, \(*ii*\) hypercube construction, \(*iii*\) dataset augmentation via relational operations, \(*iv*\) fact generation through structured comparison, and \(*v*\) interestingness filtering via maximum\-entropy selection, detailed in Algorithm[B\.3](https://arxiv.org/html/2607.15544#A2.SS3), Appendix\. By separating structured numerical reasoning from natural language generation, our framework reduces hallucination, enforces valid comparisons, and preserves quantitative consistency\.

#### Step 0\. Style Extraction\.

To ensure the generated summaries maintain stylistic continuity with domain\-specific standards, we employ a preliminary style alignment phase \(referred to here as Step 0\)\. In this phase, the LLM is provided with summaries from previous reporting rounds to extract a latent representation of the "Epidemiological Expert" tone\.

The primary objective of this extraction is to align the linguistic distribution of the key takeaways with the ground truth\. This alignment is critical for downstream evaluation; standard lexical metrics such as ROUGE and BLEU often penalize summaries that are factually and numerically accurate but stylistically divergent\. By normalizing the output style before fact generation and comparison extraction, we minimize variance attributable to phrasing and improve the sensitivity of our automated evaluation metrics\.

#### Step 1\. Representation as a Partial Order\.

We represent the dataset as a relation𝒜\(ai:𝒱i,i=1,…,K\)\\mathcal\{A\}\(a\_\{i\}:\\mathcal\{V\}\_\{i\},i=1,\\ldots,K\), whereaia\_\{i\}represents an attribute and𝒱i\\mathcal\{V\}\_\{i\}represents its domain;𝐭∈𝒱1×…​𝒱K\\mathbf\{t\}\\in\\mathcal\{V\}\_\{1\}\\times\\ldots\\mathcal\{V\}\_\{K\}correspond to possible type values\. Rather than assuming a fixed schema, we induce the underlying factor structure directly from free\-text scenario descriptions using an LLMℳ\\mathcal\{M\}\.

In the public health dataset we consider, attributes include interventions \(e\.g\., vaccination policies and compliance\), demographics \(e\.g\., age groups\), regions \(e\.g\., states\), and specific metrics \(e\.g\., infection and hospitalization rates\)\. We sometimes refer to type values𝐭\\mathbf\{t\}as scenarios and attributesaia\_\{i\}’s as scenario axes\. We use𝐭i\\mathbf\{t\}\_\{i\}to refer to the value of theiith scenario axis\.

![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/partialOrder.png)Figure 3:Partial order over attributes\.There can be a partial order \(or poset\)≤i\\leq\_\{i\}on each attributeaia\_\{i\}\. For instance, the intervention attribute for public health can consist of specific vaccination policies \(e\.g\., vaccination rates for different demographics\) and timing \(when the policy is implemented\)\. The natural hierarchy among these intervention types forms a partial order \(Figure[3](https://arxiv.org/html/2607.15544#S4.F3)\)\. These might additionally be classified as “primary” \(specifying which policy or strategy is applied\) and “modifier” \(specifying its implementation, e\.g\., timing or intensity\)\. Similarly, the hierarchy associated with demographics forms a poset\. The posets for different attributes can be combined to form a poset on the relation𝒜\\mathcal\{A\}\.

In our pipeline, this structure is constructed and exploited by two agents\. A*decomposition agent*ℳ𝒟\\mathcal\{M\_\{D\}\}maps each free\-text scenario description to a structured tuple𝐭∈𝒱1×⋯×𝒱K\\mathbf\{t\}\\in\\mathcal\{V\}\_\{1\}\\times\\cdots\\times\\mathcal\{V\}\_\{K\}, distinguishing primary from modifier axes\. A*comparison\-selection agent*then traverses the induced product poset to identify three classes of structurally valid comparisons — defined formally in Step 3\(§[4](https://arxiv.org/html/2607.15544#S4.SS0.SSS0.Px4)\)\.

Robust schema inference\.Data formats are not fully consistent and uniform\. We use an agentic component to learn a poset representation for the dataset\. The model is provided with a statistical profile of the dataset–field names, data types, cardinalities, and sample values–to produce a canonical internal representationσ\\sigma\. If schema loading fails or raises a validation errorϵ\(t\)\\epsilon^\{\(t\)\}, the representation is corrected iteratively:σ\(t\+1\)←ℳ​\(σ\(t\),ϵ\(t\)\),\\small\\sigma^\{\(t\+1\)\}\\;\\leftarrow\\;\\mathcal\{M\}\\\!\\left\(\\sigma^\{\(t\)\},\\,\\epsilon^\{\(t\)\}\\right\), up to a maximum ofTTattempts\. This allows automatic diagnosis and correction of type mismatches or incorrect column identifications without manual intervention\. Let𝒟\\mathcal\{D\}denote the learned dataset\. For each datapoint𝐱∈𝒟\\mathbf\{x\}\\in\\mathcal\{D\}, let𝐭​\(𝐱\)\\mathbf\{t\}\(\\mathbf\{x\}\)denote the corresponding type descriptor\.

#### Step 2\. Dataset Augmentation\.

Starting from𝒟\\mathcal\{D\}, we construct𝒟aug\\mathcal\{D\}\_\{\\mathrm\{aug\}\}using a sequence of operations, and augment it via deterministic relational operations\. Examples of operations are*unions*\. A union of a subsetS⊂𝒟a​u​gS\\subset\\mathcal\{D\}\_\{aug\}can be done if there exists an attributeiisuch that: \(1\) for allj≠ij\\neq i,𝐭​\(𝐱\)j,𝐱∈S\\mathbf\{t\}\(\\mathbf\{x\}\)\_\{j\},\\mathbf\{x\}\\in Sare the same, and \(2\) there exists an itemy∈𝒱iy\\in\\mathcal\{V\}\_\{i\}which is a parent of𝐭​\(𝐱\)i,𝐱∈S\\mathbf\{t\}\(\\mathbf\{x\}\)\_\{i\},\\mathbf\{x\}\\in S\. Data points corresponding to such a subsetSScan be combined to form an additional data point𝐱​\(S\)\\mathbf\{x\}\(S\)with a value that depends on how they can be aggregated; we set the associated type𝐭​\(𝐱\)i=y\\mathbf\{t\}\(\\mathbf\{x\}\)\_\{i\}=y\. For instance, the value in𝐱​\(S\)\\mathbf\{x\}\(S\)might be a sum of the values of the individual items, or can correspond to the associated range \(between the minimum to the maximum\)\. An agent is implemented to apply the grammar rules and combine such subsets of items \(with chain of thought reasoning to ensure correctness\), which are added repeatedly to the dataset𝒟a​u​g\\mathcal\{D\}\_\{aug\}\.

#### Step 3\. Fact Generation\.

We now specify a grammar for generating facts from𝒟aug\\mathcal\{D\}\_\{\\mathrm\{aug\}\}\. This is done in two ways\. First, each𝐱∈𝒟a​u​g\\mathbf\{x\}\\in\\mathcal\{D\}\_\{aug\}can be translated to a fact\. Second, we generate a set of facts with quantitative claims about relationships between scenarios, using a*Comparison Grammar*that restricts which pairs are valid, preventing spurious “apples\-to\-oranges” analyses\. It helps the LLM by providing a structured guideline for generating the summary, instead of relying entirely on the LLM’s random or unconstrained generation\.

Examples of rules satisfied by valid comparisons in the grammar are: \(R1\) two scenarios differ on*exactly one*axis while agreeing on all others; \(R2\) the baseline scenario is compared against a pool of active\-condition scenarios that all share at least one common secondary axis value; or \(R3\) two such pools are compared only when their members are matched on every axis except the one of interest\. Each comparison generated by the grammar is associated with a quantity such as the percentage change in mean summary statistic between the two sides\. We use𝒞\\mathcal\{C\}to denote the set of all outputs generated by the comparison grammar; these are translated to facts in the narrative\. This ensures that every fact is numerically grounded and tied to a specific attribute\.

#### Step 4\. Generate interesting facts\.

The set𝒞\\mathcal\{C\}of facts may contain redundant or low\-information comparisons\. There are many ways in which a statement can become interesting, e\.g\., changes in temporal and spatial patterns, or relative to reports in the literature\. We formalize interestingness selection using a maximum\-entropy \(MaxEnt\) approachJaynes\([1957](https://arxiv.org/html/2607.15544#bib.bib91)\)and construct a compact, high\-salience subset𝒞∗⊆𝒞\\mathcal\{C\}^\{\*\}\\subseteq\\mathcal\{C\}\. Each comparisonc=\(A,B\)∈𝒞c=\(A,B\)\\in\\mathcal\{C\}is associated with a*feature profile*ϕ​\(c\)∈ℱ\\phi\(c\)\\in\\mathcal\{F\}that records which combination of scenario axes, target quantity, subgroup, spatial unit, and any other indexed dimension it addresses\. These are computed using Iterative Proportional Fitting \(IPF\), which alternately normalizes the distribution to satisfy each dimension’s constraint in turn and converges to the maximum\-entropy solution\. This discourages redundancy \(e\.g\., multiple comparisons addressing the same target or subgroup\) while rewarding facts that illuminate underrepresented cells of𝒞\\mathcal\{C\}\.

For state\-level interestingness, we instantiate this selection problem over a state\-by\-field observation matrix, where each cell represents a normalized projection statistic for a state and epidemiological field\. Let𝐌∈ℝN×P\\mathbf\{M\}\\in\\mathbb\{R\}^\{N\\times P\}be an observation matrix whereMi​jM\_\{ij\}is the point\-estimate projection for stateiion epidemiological fieldjj\(NNU\.S\. states,PPfields from SMH Round\)\. Fields span heterogeneous units, so we column\-normalise:M~i​j=Mi​j−μjσj\\tilde\{M\}\_\{ij\}=\\frac\{M\_\{ij\}\-\\mu\_\{j\}\}\{\\sigma\_\{j\}\} The goal is to identify cells\(i,j\)\(i,j\)whose values are surprising and not explained by state\-level effects alone\.We evaluate the model for the expected valueFi​jF\_\{ij\}, against which the residualRi​j=M~i​j−Fi​jR\_\{ij\}=\\tilde\{M\}\_\{ij\}\-F\_\{ij\}defines surprise\.

#### MaxEnt/IPF\.

We fit a maximum\-entropy matrix𝐅\\mathbf\{F\}using Iterative Proportional Fitting \(IPF\), starting from a uniform prior and constraining it to match the observed marginal totals for all regions, field categories, and outcome metrics\. To penalise redundant findings, we measure the entropy of the state distribution in the top\-KKlist\. A higherHKH\_\{K\}indicates findings spread across more states\.HK=−∑ip^i​log⁡p^iH\_\{K\}=\-\\sum\_\{i\}\\hat\{p\}\_\{i\}\\log\\hat\{p\}\_\{i\}

#### Final Summary Generation\.

Once the top surprising cells are identified, we generate a written summary using a large language model \(LLM\)\. Rather than prompting the LLM directly with raw numbers, we use a three\-step chain\-of\-thought \(CoT\) pipeline to ensure structured reasoning\. Step A: ObserveThe LLM is given a structured findings block containing the top\-KKsurprising\(i,j\)\(i,j\)cells, their raw values, residuals, and regional aggregates\. It is asked to describe what is driving each state’s surprise score and identify any regional or category\-level patterns\. Step B: ReflectThe LLM is given its Step A output and asked to select the 2–3 most policy\-relevant findings\. For each, it produces a one\-sentence finding, an explanation of why the deviation is epidemiologically meaningful\. Step C: SynthesizeUsing the observations from Steps A and B, the LLM writes a concise national executive summary in plain language, suitable for a public health audience\.

The three\-step structure serves two purposes\. First, it separates pattern detection \(Step A\) from interpretation \(Step B\) from communication \(Step C\), reducing the chance that the LLM fabricates explanations for statistical artifacts\. Second, each step’s output is logged alongside its prompt, making the reasoning chain understandable\.

## 5Experimental Setup

Data\.We evaluate our framework on theScenario Modeling Hub \(SMH\)dataset \(referred to as GT\), a multi\-round ensemble modeling initiative for infectious disease projections\. We specifically focus on COVID\-19 projections, where each modeling round poses a structured set of scenarios defined along multiple axes: e\.g\., booster coverage \(none, 65\+, all ages\), immune escape \(low, high\), and timing \(classic, early\), resulting in 4–6 distinct scenarios per round\. Each scenario is projected by multiple independent modeling teams, which are then ensembled to produce quantile outputs over the projection horizon\.

Dataset characteristics\.*Scenario descriptions*: Free\-text markdown files \(150–400 words per scenario\) on modeling assumptions, booster coverage and immune escape parameters \( for the most recent round\), and related epidemiological context\. *Ensemble projections*: Probabilistic \(quantile\-based\) outputs per scenario across target quantities \(hospitalizations, deaths\), stratified by age group \(\|𝒢\|=2\|\\mathcal\{G\}\|=2: 0–64, 65\+\) and jurisdiction \(\|ℛ\|=51\|\\mathcal\{R\}\|=51U\.S\. states \+ national\)\.*Reference summaries\.*Expert\-curated national executive summaries \(400–800 words\) as well as templatized state\-level summaries, serving as ground truth for evaluation\.

Implementation Details\.All LLM calls useGemma 4 26B\(gemma\-4\-26b\-a4b\-it\)Google DeepMind\([2026](https://arxiv.org/html/2607.15544#bib.bib92)\)via the Google GenAI API free tier, with temperature 0\.0 for reasoning and 0\.2–0\.4 for generation; the state\-level MaxEnt summary usesGemma 4 31B, no fine\-tuning is performed\. For interestingness analysis, we use ipfn package to implement IPF; M3 metrics use scikit\-learn, and embedding similarity uses sentence\-transformers \(all\-MiniLM\-L6\-v2\)Reimers and Gurevych\([2019](https://arxiv.org/html/2607.15544#bib.bib93)\)\. All prompts and responses are logged to JSONL for auditability\. The full pipeline runs in 3–5 minutes per round, dominated by API latency\.

## 6Results

We report our results across two distinct dimensions\. While NLP metrics \(BLEU, ROUGE\) measure adherence to the existing GT, they may paradoxically penalize our framework for identifying "MaxEnt" \(Maximum Entropy\) insights—critical epidemiological observations that are statistically present in the data but were not reported in the GT\.

Because the GT is not an exhaustive gold standard, a summary that captures higher\-utility insights may receive lower lexical overlap scores\. Consequently, we separate our findings intoLexical Alignment \(closeness to GT\)and Expert Utility \(human ranking of "interestingness"\)\. We demonstrate that our framework consistently surfaces important components that experts prefer, even when those components are absent from the manual reports\.

### 6\.1Evaluation Metrics

We evaluate generated summaries using three complementary metrics: factual fidelity, coverage of ground\-truth findings, and stylistic consistency with the reference report\.

Table 1:Comparison of M1, M2, and M3 scores across SMH rounds\. For M1, “100%/30” denotes 100% factual fidelity over 30 extracted numerical claims\.Factual accuracy \(M1\)\.M1 measures whether the generated summary makes numerically correct claims\. We use a feedback LLM to extract quantitative assertions from the generated summary and verify each claim against the source projection tables\. We also check whether a number is directly reported or derived, since values such as percentage increases or decreases may not appear explicitly in the table\. The final score is the fraction of verified numerical claims:M1=NverifiedNclaims\.\\mathrm\{M1\}=\\frac\{N\_\{\\mathrm\{verified\}\}\}\{N\_\{\\mathrm\{claims\}\}\}\.

*Caveat\.*Since M1 is based on LLM\-extracted numerical claims and LLM\-based verification, the score can be stochastic and imperfect\. In some cases, a correctly derived value may be marked as incorrect if the verifier fails to trace the arithmetic or match the derived number to the source tables \(Appendix[7](https://arxiv.org/html/2607.15544#A2.F7)\)\. Therefore, M1 should be interpreted as an approximate factual\-fidelity measure rather than an exact ground\-truth score\.

Another limitation is that M1 does not directly capture the amount of quantitative evidence included in the summary \(Table[1](https://arxiv.org/html/2607.15544#S6.T1)\)\. For example, a summary that contains only one numerical claim can receive a perfect M1 score if that claim is correct, even though it may be less informative than a number\-rich summary with broader quantitative coverage\.

Coverage \(M2\)\.M2 measures how well the generated summary captures the main scenario\-based findings from the ground\-truth \(GT\) Key Takeaways\. The evaluator first extracts distinct GT information units, such as intervention effects, marginal gains, absolute burden under named scenarios, and age\-group burden patterns\. Each GT point is then marked as covered, partially covered, or missing in the generated summary\. We then compute:

M2=Ncovered\+12​NpartialNGT\.\\small\\mathrm\{M2\}=\\frac\{N\_\{\\mathrm\{covered\}\}\+\\frac\{1\}\{2\}N\_\{\\mathrm\{partial\}\}\}\{N\_\{\\mathrm\{GT\}\}\}\.Style consistency \(M3\)\.M3 measures whether the generated summary follows the writing style of the GT Key Takeaways\. We compare the generated and GT summaries using cosine similarity over their text representations\. In implementation, this is computed primarily using embedding similarity\.

M3=cos⁡\(𝐞gen,𝐞GT\)=𝐞gen⊤​𝐞GT‖𝐞gen‖2​‖𝐞GT‖2\.\\small\\mathrm\{M3\}=\\cos\\\!\\left\(\\mathbf\{e\}\_\{\\mathrm\{gen\}\},\\mathbf\{e\}\_\{\\mathrm\{GT\}\}\\right\)=\\frac\{\\mathbf\{e\}\_\{\\mathrm\{gen\}\}^\{\\top\}\\mathbf\{e\}\_\{\\mathrm\{GT\}\}\}\{\\left\\lVert\\mathbf\{e\}\_\{\\mathrm\{gen\}\}\\right\\rVert\_\{2\}\\left\\lVert\\mathbf\{e\}\_\{\\mathrm\{GT\}\}\\right\\rVert\_\{2\}\}\.
Table[1](https://arxiv.org/html/2607.15544#S6.T1)shows thatEpiNarrateachieves strong M1 scores across rounds while making a larger number of numerical claims, which is desirable for quantitative public\-health summaries\.EpiNarratealso achieves the highest M2 coverage, suggesting that the partial\-order representation helps recover more valid scenario\-axis comparisons than unconstrained baselines\. For M3,EpiNarrateremains comparable while providing stronger quantitative grounding and coverage\.

### 6\.2MaxEnt\-based Interestingness Analysis

To identify state\-level patterns that are not apparent from aggregate national summaries alone, we apply a maximum\-entropy interestingness analysis to the state\-level projection outputs\. The MaxEnt baseline is fitted using iterative proportional fitting \(IPF\), simultaneously matching HHS regional marginals \(row axis\) and field\-category/metric marginals \(column axis\)\. All absolute burden and averted\-count fields are expressed per 100,000 population using 2022 ACS five\-year estimates before the matrix is constructed, so that state size does not mechanically inflate residual scores\. Large absolute residuals indicate state–metric cells that are surprising relative to what would be expected from the marginal structure alone\. We assign states to HHS regions, which group U\.S\. states and territories into ten regions[U\.S\. Department of Health and Human Services](https://arxiv.org/html/2607.15544#bib.bib86)\.

![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/maxEnt.png)Figure 4:MaxEnt\-ranked residual cells for R19 with human interestingness ratings\.*Evaluation and justification\.*Figure[4](https://arxiv.org/html/2607.15544#S6.F4)shows the top MaxEnt cells with human interestingness ratings \(Details in Appendix Fig,[12](https://arxiv.org/html/2607.15544#A3.F12),[13](https://arxiv.org/html/2607.15544#A3.F13),[14](https://arxiv.org/html/2607.15544#A3.F14),\)\. High positive residuals highlight above\-expected values, while negative residuals indicate below\-expected values\. Several high\-ranked cells involve age\-specific mortality shares: Alaska, New Mexico, and Mississippi have higher\-than\-expected shares of projected deaths among adults aged 65\+\. Other high\-ranked findings capture elevated hospitalization burden in West Virginia, Kentucky, and New York, as well as stronger\-than\-expected intervention effects on averted hospitalizations in Vermont, New Hampshire, and Maine\. DC, Rhode Island, and Utah show lower\-than\-expected death burden\.

Expert ratings suggest that MaxEnt can be useful for state\-level signals, but not all statistically surprising cells are equally narrative\-worthy\. Age\-share findings generally receive higher ratings, while some hospitalization\-burden findings are rated lower because they may require additional context or may be less policy\-informative\. Overall, the MaxEnt analysis provides a complementary selection mechanism for identifying state\-level deviations, while human grading helps distinguish statistically surprising cells from genuinely useful narrative content\.

### 6\.3Baselines

We compareEpiNarrateagainst five families spanning the data\-to\-text design space, along with a schema\-only information\-deprived floor\. All baselines are evaluated under the same protocol on rounds R17, R18, and R19\. Table[2](https://arxiv.org/html/2607.15544#S6.T2)summarizes the baseline categories, implementation details, and main observed limitations\.

Table 2:Summary of baseline families, implementation choices, and observed limitations\.Overall, decoder\-based LLM baselines often produce fluent summaries and can achieve high surface\-level alignment scores, but they are less reliable in factual fidelity and coverage because numerical reasoning and comparison selection are not explicitly constrained\. CoT prompting improves some arithmetic reasoning, but without a structured scenario representation, it can still miss valid scenario\-axis comparisons or reason over invalid contrasts\. Encoder–decoder baselines such as T5\-Large and BART tend to produce shorter and more generic summaries with very few numerical claims; this can sometimes inflateM1M\_\{1\}when the claim count is small, but generally leads to poor coverage of scenario\-based ground\-truth findings\. In contrast,EpiNarratemaintains a larger number of grounded numerical claims while achieving the strongest coverage across rounds\.

### 6\.4Ablation Study

To evaluate the contribution of each component inEpiNarrate, we conduct an ablation study by removing one module at a time while keeping the rest of the pipeline unchanged\. We focus on three core components:1\.chain\-of\-thought reasoning,2\.partial order, and3\.preprocessing/schema agent\. The goal is to assess whether each module contributes to factual fidelity, coverage, and alignment with expert\-written summaries\.

Table 3:Ablation study ofEpiNarratecomponentsThe ablation results in Table[3](https://arxiv.org/html/2607.15544#S6.T3)show that CoT mainly improves factual fidelity by helping the model trace and verify derived numerical claims, increasing M1 from 70% to 100%\. In contrast, the partial\-order representation is most important for coverage: without it, the model still produces numerically correct claims, but misses many valid scenario\-axis comparisons, causing M2 to drop from 0\.78 to 0\.28\.*EpiNarratew/o Preprocessing Agent*cannot be applied robustly across different SMH rounds, as it leads to more schema and type\-mismatch errors\. This component is necessary in our pipeline because each round may have different ensemble data formats, scenario definitions, and intervention policies\.

## 7Conclusion

We introducedEpiNarrate, an agentic framework for generating grounded and comprehensive narratives from epidemiological scenario\-modeling outputs\. Rather than prompting an LLM directly over large projection tables, our approach separates structured numerical reasoning from natural\-language generation\. The framework first decomposes scenario descriptions into structured axes, builds an augmented scenario space, generates valid comparisons through a comparison grammar, and then selects salient findings using interestingness\-based filtering, including MaxEnt/IPF\. Across SMH rounds,EpiNarrateimproves coverage of key scenario\-based findings while maintaining strong factual fidelity of numerical claims\. The results suggest that explicit structure is important for this task\. The MaxEnt analysis further shows that the framework can surface state\-level and age\-specific patterns that may not appear in aggregate national summaries, providing an additional mechanism for identifying potentially policy\-relevant findings\. Limitations and Future Work\.Some statistically surprising MaxEnt signals may require human review to distinguish meaningful patterns from noise; however, this human\-in\-the\-loop step can also strengthen the reliability and usefulness of the generated public\-health narratives\. Our evaluation is currently limited to scenario\-modeling rounds with factorial designs, and the pipeline uses Gemma 4 models accessed through a free\-tier API\. Performance may therefore vary across tasks, scenario structures, or model backbones\. ValidatingEpiNarrateacross more diverse public\-health settings and LLMs remains an important direction for future work\.

## References

- \[1\]J\. Achiam, S\. Adler, S\. Agarwal, L\. Ahmad, I\. Akkaya, F\. L\. Aleman, D\. Almeida, J\. Altenschmidt, S\. Altman, S\. Anadkat,et al\.\(2023\)Gpt\-4 technical report\.arXiv preprint arXiv:2303\.08774\.Cited by:[Table 2](https://arxiv.org/html/2607.15544#S6.T2.7.7.7.7.1.1)\.
- \[2\]\(2026\)Introducing Claude Opus 4\.6\.Note:[https://www\.anthropic\.com/news/claude\-opus\-4\-6](https://www.anthropic.com/news/claude-opus-4-6)Accessed: 2026\-05\-07Cited by:[Table 2](https://arxiv.org/html/2607.15544#S6.T2.7.7.7.7.1.1)\.
- \[3\]V\. Balachandran, A\. Pagnoni, J\. Y\. Lee, D\. Rajagopal, J\. Carbonell, and Y\. Tsvetkov\(2021\)StructSum: summarization via structured representations\.InProceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics \(EACL\),pp\. 2575–2585\.External Links:[Link](https://aclanthology.org/2021.eacl-main.220/)Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p4.1)\.
- \[4\]CDCForecasts of Flu Hospital Admissions\.Note:[https://www\.cdc\.gov/flu\-forecasting/data\-vis/current\-week\.html](https://www.cdc.gov/flu-forecasting/data-vis/current-week.html)Cited by:[§1](https://arxiv.org/html/2607.15544#S1.p1.1)\.
- \[5\]CDCMMWR Surveillance Summaries\.Note:[https://www\.cdc\.gov/mmwr/indss\_2025\.html](https://www.cdc.gov/mmwr/indss_2025.html)Cited by:[§1](https://arxiv.org/html/2607.15544#S1.p1.1)\.
- \[6\]CDCScenario Modeling Hub\.Note:[https://covid19scenariomodelinghub\.org/](https://covid19scenariomodelinghub.org/)Cited by:[§1](https://arxiv.org/html/2607.15544#S1.p1.1)\.
- \[7\]W\. Chen\(2023\)Large language models are few\(1\)\-shot table reasoners\.InFindings of the Association for Computational Linguistics: EACL,pp\. 1090–1100\.External Links:[Link](https://aclanthology.org/2023.findings-eacl.83/)Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p3.1)\.
- \[8\]V\. Dibia\(2023\)LIDA: a tool for automatic generation of grammar\-agnostic visualizations and infographics using large language models\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics: System Demonstrations \(ACL\),pp\. 113–126\.External Links:[Link](https://arxiv.org/abs/2303.02927)Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p3.1),[Table 2](https://arxiv.org/html/2607.15544#S6.T2.1.1.1.2.1.1)\.
- \[9\]T\. C\. Ferreira, C\. van der Lee, E\. Van Miltenburg, and E\. Krahmer\(2019\)Neural data\-to\-text generation: a comparison between pipeline and end\-to\-end architectures\.InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th International joint conference on natural language processing \(EMNLP\-IJCNLP\),pp\. 552–562\.Cited by:[§1](https://arxiv.org/html/2607.15544#S1.p2.1),[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p1.1),[Table 2](https://arxiv.org/html/2607.15544#S6.T2.10.10.10.2.1.1)\.
- \[10\]Google DeepMind\(2026\)Gemma 4\.Note:[https://deepmind\.google/models/gemma/gemma\-4/](https://deepmind.google/models/gemma/gemma-4/)Accessed: 2026\-05\-07Cited by:[§5](https://arxiv.org/html/2607.15544#S5.p3.1)\.
- \[11\]E\. T\. Jaynes\(1957\)Information theory and statistical mechanics\.Physical review106\(4\),pp\. 620\.Cited by:[§4](https://arxiv.org/html/2607.15544#S4.SS0.SSS0.Px5.p1.5)\.
- \[12\]T\. Khot, H\. Trivedi, M\. Finlayson, Y\. Fu, K\. Richardson, P\. Clark, and A\. Sabharwal\(2023\)Decomposed prompting: a modular approach for solving complex tasks\.InThe Eleventh International Conference on Learning Representations \(ICLR\),External Links:[Link](https://openreview.net/forum?id=_nGgzQjzaRy)Cited by:[Table 2](https://arxiv.org/html/2607.15544#S6.T2.9.9.9.3.1.1)\.
- \[13\]R\. Koncel\-Kedziorski, D\. Bekal, Y\. Luan, M\. Lapata, and H\. Hajishirzi\(2019\)Text generation from knowledge graphs with graph transformers\.InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics \(NAACL\),pp\. 2284–2293\.External Links:[Link](https://aclanthology.org/N19-1238/)Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p4.1)\.
- \[14\]R\. Lebret, D\. Grangier, and M\. Auli\(2016\)Neural text generation from structured data with application to the biography domain\.InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 1203–1213\.External Links:[Link](https://aclanthology.org/D16-1128/)Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p1.1)\.
- \[15\]M\. Lewis, Y\. Liu, N\. Goyal, M\. Ghazvininejad, A\. Mohamed, O\. Levy, V\. Stoyanov, and L\. Zettlemoyer\(2020\)BART: denoising sequence\-to\-sequence pre\-training for natural language generation, translation, and comprehension\.InProceedings of the 58th annual meeting of the association for computational linguistics,pp\. 7871–7880\.Cited by:[Table 2](https://arxiv.org/html/2607.15544#S6.T2.10.10.10.2.1.1)\.
- \[16\]MIDAS Network\(2025\)COVID\-19 scenario modeling hub\.Note:[https://github\.com/midas\-network/covid19\-scenario\-modeling\-hub](https://github.com/midas-network/covid19-scenario-modeling-hub)Accessed: 2026\-05\-01Cited by:[§1](https://arxiv.org/html/2607.15544#S1.p1.1)\.
- \[17\]OpenAI\(2026\)GPT\-5\.5\.Note:[https://developers\.openai\.com/api/docs/models/gpt\-5\.5](https://developers.openai.com/api/docs/models/gpt-5.5)Accessed: 2026\-05\-07Cited by:[Table 2](https://arxiv.org/html/2607.15544#S6.T2.7.7.7.7.1.1)\.
- \[18\]A\. Parikh, X\. Wang, S\. Gehrmann, M\. Faruqui, B\. Dhingra, D\. Yang, and D\. Das\(2020\)ToTTo: a controlled table\-to\-text generation dataset\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 1173–1186\.Cited by:[§1](https://arxiv.org/html/2607.15544#S1.p2.1),[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p1.1)\.
- \[19\]R\. Puduppully, L\. Dong, and M\. Lapata\(2019\)Data\-to\-text generation with content selection and planning\.InProceedings of the AAAI conference on artificial intelligence,Vol\.33,pp\. 6908–6915\.Cited by:[§1](https://arxiv.org/html/2607.15544#S1.p2.1),[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p2.1),[Table 2](https://arxiv.org/html/2607.15544#S6.T2.10.10.12.1.1.1.1)\.
- \[20\]R\. Puduppully, L\. Dong, and M\. Lapata\(2019\)Data\-to\-text generation with entity modeling\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics \(ACL\),pp\. 2023–2035\.External Links:[Link](https://aclanthology.org/P19-1195/)Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p2.1)\.
- \[21\]C\. Raffel, N\. Shazeer, A\. Roberts, K\. Lee, S\. Narang, M\. Matena, Y\. Zhou, W\. Li, and P\. J\. Liu\(2020\)Exploring the limits of transfer learning with a unified text\-to\-text transformer\.Journal of machine learning research21\(140\),pp\. 1–67\.Cited by:[Table 2](https://arxiv.org/html/2607.15544#S6.T2.10.10.10.2.1.1)\.
- \[22\]C\. Rebuffel, L\. Soulier, G\. Scoutheeten, and P\. Gallinari\(2020\)A hierarchical model for data\-to\-text generation\.InProceedings of the European Conference on Information Retrieval \(ECIR\),pp\. 65–80\.External Links:[Link](https://arxiv.org/abs/1912.10011)Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p2.1)\.
- \[23\]N\. G\. Reich, J\. Lessler, S\. Funk, C\. Viboud, A\. Vespignani, R\. J\. Tibshirani, K\. Shea, M\. Schienle, M\. C\. Runge, R\. Rosenfeld,et al\.\(2022\)Collaborative hubs: making the most of predictive epidemic modeling\.Vol\.112,American Public Health Association\.Cited by:[§1](https://arxiv.org/html/2607.15544#S1.p1.1)\.
- \[24\]N\. Reimers and I\. Gurevych\(2019\)Sentence\-bert: sentence embeddings using siamese bert\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing,Cited by:[§5](https://arxiv.org/html/2607.15544#S5.p3.1)\.
- \[25\]E\. Reiter and R\. Dale\(2000\)Building natural language generation systems\.Cambridge University Press\.Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p1.1)\.
- \[26\]Y\. Sui, M\. Zhou, M\. Zhou, S\. Han, and D\. Zhang\(2024\)Table meets LLM: can large language models understand structured table data? an empirical study on effective serialization\.InProceedings of the 17th ACM International Conference on Web Search and Data Mining \(WSDM\),pp\. 645–654\.External Links:[Link](https://arxiv.org/abs/2305.13062)Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p3.1)\.
- \[27\]U\.S\. Department of Health and Human ServicesRegional offices\.Note:[https://www\.hhs\.gov/about/agencies/iea/regional\-offices/index\.html](https://www.hhs.gov/about/agencies/iea/regional-offices/index.html)Accessed: 2026\-05\-06Cited by:[§6\.2](https://arxiv.org/html/2607.15544#S6.SS2.p1.1)\.
- \[28\]X\. Wang, J\. Wei, D\. Schuurmans, Q\. V\. Le, E\. H\. Chi, S\. Narang, A\. Chowdhery, and D\. Zhou\(2023\)Self\-consistency improves chain of thought reasoning in language models\.InThe Eleventh International Conference on Learning Representations \(ICLR\),External Links:[Link](https://openreview.net/forum?id=1PL1NIMMrw)Cited by:[Table 2](https://arxiv.org/html/2607.15544#S6.T2.9.9.9.3.1.1)\.
- \[29\]J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, B\. Ichter, F\. Xia, E\. H\. Chi, Q\. V\. Le, and D\. Zhou\(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.InProceedings of the 36th International Conference on Neural Information Processing Systems \(NeurIPS\),External Links:[Link](https://arxiv.org/abs/2201.11903)Cited by:[Table 2](https://arxiv.org/html/2607.15544#S6.T2.9.9.9.3.1.1)\.
- \[30\]WHODisease Outbreak News \(DONs\)\.Note:[https://www\.who\.int/emergencies/disease\-outbreak\-news](https://www.who.int/emergencies/disease-outbreak-news)Cited by:[§1](https://arxiv.org/html/2607.15544#S1.p1.1)\.
- \[31\]S\. Wiseman, S\. M\. Shieber, and A\. M\. Rush\(2017\)Challenges in data\-to\-document generation\.InProceedings of the 2017 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 2253–2263\.External Links:[Link](https://aclanthology.org/D17-1239/)Cited by:[§2](https://arxiv.org/html/2607.15544#S2.SS0.SSS0.Px1.p1.1)\.
- \[32\]D\. Zhou, N\. Schärli, L\. Hou, J\. Wei, N\. Scales, X\. Wang, D\. Schuurmans, C\. Cui, O\. Bousquet, Q\. V\. Le, and E\. H\. Chi\(2023\)Least\-to\-most prompting enables complex reasoning in large language models\.InThe Eleventh International Conference on Learning Representations \(ICLR\),External Links:[Link](https://openreview.net/forum?id=WZH7099tgfM)Cited by:[Table 2](https://arxiv.org/html/2607.15544#S6.T2.9.9.9.3.1.1)\.

## Appendix AAdditional Details on Partial Ordering

The partial order is the central structural device inEpiNarrate\. It serves three roles: \(i\) guiding the LLM decomposition agent toward a consistent axis representation, \(ii\) defining which scenario subsets can be legitimately merged during augmentation, and \(iii\) constraining the comparison grammar so that every emitted fact corresponds to a valid, interpretable contrast\.

#### Axis extraction \(agentic\)\.

Each round contains free\-text scenario descriptions that implicitly encode the experimental design: which factors vary, which are held fixed, and which scenario serves as the no\-intervention baseline\. Algorithm[2](https://arxiv.org/html/2607.15544#alg2)recovers this structure by promptingℳ\\mathcal\{M\}to map each scenario to a tuple of axis values and to classify each axis asprimary\(the intervention itself, e\.g\. vaccination target population\) ormodifier\(a parameter of the intervention, e\.g\. timing\)\. The distinction induces a two\-level partial order on the intervention attribute: primary values are incomparable to each other \(they represent qualitatively different strategies\), while modifier values refine a given primary choice and are therefore ordered beneath it\. This hierarchy is what prevents the comparison grammar from pooling across primary axes in grouped\-counterfactual comparisons \(Rule R2\)\.

#### Dataset augmentation\.

Given the per\-scenario axis tuples, the augmentation step \(Algorithm[4](https://arxiv.org/html/2607.15544#alg4)\) constructs𝒟aug\\mathcal\{D\}\_\{\\mathrm\{aug\}\}by taking all2\|𝒮\|−12^\{\|\\mathcal\{S\}\|\}\{\-\}1non\-empty subsets of the scenario set\. A union of subsetCCis valid when there exists an attributeiisuch that all members ofCCagree on every other attributej≠ij\\neq iand the value on attributeiihas a common parent in the partial order\. In practice, the union table records—for each subset—which axis values are shared \(common axes, the intersection⋂s∈C𝐭s\\bigcap\_\{s\\in C\}\\mathbf\{t\}\_\{s\}\) and which vary \(union axes,⋃s∈C𝐭s\\bigcup\_\{s\\in C\}\\mathbf\{t\}\_\{s\}\)\. Per\-target statistics \(min, max, mean of member medians\) are attached for every \(age group, target\) pair\. Figure[5](https://arxiv.org/html/2607.15544#A1.F5)shows an excerpt: singleton rows correspond to individual scenarios; multi\-member rows aggregate statistics across the subset, enabling downstream comparisons at different levels of granularity\.

#### Filtering with grammar rules\.

The comparison grammar \(Algorithm[5](https://arxiv.org/html/2607.15544#alg5)\) traverses𝒟aug\\mathcal\{D\}\_\{\\mathrm\{aug\}\}and emits only structurally valid comparisons\. Rule R1 isolates marginal effects by requiring exactly one axis of difference\. Rule R2 compares the counterfactual against a pooled intervention group, with the guard that within\-group variation must be confined to modifier axes—without this constraint, the pool would mix fundamentally different interventions \(e\.g\. vaccinating high\-risk only vs\. all ages\), making causal attribution impossible\. Rule R3 permits pool\-vs\-pool comparisons only when the two sides are balanced on all non\-varying axes\. Figure[6](https://arxiv.org/html/2607.15544#A1.F6)shows the resulting filtered table: each row is a grammar\-valid comparison annotated with its type, the varying axis, and the percentage change in each outcome\.

![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/augmented.png)Figure 5:Excerpt from the augmented dataset𝒟aug\\mathcal\{D\}\_\{\\mathrm\{aug\}\}for Round 19\. Rows withcombo\_size=1\\,\{=\}\\,1are individual scenarios; larger subsets aggregate member\-level median projections\. Thecommon\_\_\*columns show axis values shared by all members of the subset \(intersection\), whileunion\_\_\*columns list all values present \(union\)\. Numeric columns report the mean of member medians for each target, which are used by the comparison grammar to compute percentage changes\.![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/filtered.png)Figure 6:Excerpt from the filtered comparison table𝒞\\mathcal\{C\}produced by the comparison grammar \(Algorithm[5](https://arxiv.org/html/2607.15544#alg5)\)\. Each row is a structurally valid comparison: MP = marginal pair \(Rule R1, exactly one axis differs\), GC = grouped counterfactual \(Rule R2, baseline vs\. pooled intervention group\), SD = second\-degree \(Rule R3, balanced pool vs\. pool\)\. Thevarying\_axiscolumn identifies the factor being compared;pct\_changecolumns report the relative difference between groups A and B for each outcome and age group\.

## Appendix BPipeline

Algorithm[1](https://arxiv.org/html/2607.15544#alg1)extracts Key Takeaways from prior\-round GT reports and promptsℳ\\mathcal\{M\}to summarize their recurring structural and stylistic patterns into a style descriptionSS\.

Algorithm[2](https://arxiv.org/html/2607.15544#alg2)parses the round README for scenario metadata and callsℳ\\mathcal\{M\}to decompose scenarios into axes𝒜\\mathcal\{A\}with primary/modifier rolesρ\\rhoand per\-scenario values𝐭s\\mathbf\{t\}\_\{s\}\. A structural fallback \(underscore splitting\) is used if the LLM is unavailable\.

Algorithm[3](https://arxiv.org/html/2607.15544#alg3)maps ensemble data columns to canonical names via an identity seed refined byℳ\\mathcal\{M\}\. On failure, the tracebackϵ\(t\)\\epsilon^\{\(t\)\}is sent back toℳ\\mathcal\{M\}for diagnosis; the corrected mappingσ\(t\+1\)←ℳ​\(σ\(t\),ϵ\(t\)\)\\sigma^\{\(t\+1\)\}\\leftarrow\\mathcal\{M\}\(\\sigma^\{\(t\)\},\\epsilon^\{\(t\)\}\)is retried up toT=4T\{=\}4times\.

Algorithm[4](https://arxiv.org/html/2607.15544#alg4)enumerates all2\|𝒮\|−12^\{\|\\mathcal\{S\}\|\}\{\-\}1scenario subsets, computes common/union axes and per\-target statistics \(min, max, mean of member medians\) for each \(age group, target\) pair\.

Algorithm[5](https://arxiv.org/html/2607.15544#alg5)generates candidate facts𝒞\\mathcal\{C\}via three rules: R1 \(marginal pairs differing on one axis\), R2 \(counterfactual vs\. pooled group, guarded against mixing primary axes\), R3 \(balanced pool\-vs\-pool\)\. Each comparison is expanded across age groups with attachedΔ%\\Delta\\%\.

Algorithm[6](https://arxiv.org/html/2607.15544#alg6)fits a MaxEnt baselineFFvia IPF from a uniform prior\.

Surprise is then scored as a multinomial z\-score:

zi​j=Mi​j−Fi​jFi​j​\(1−pi​j\),pi​j=Fi​j∑jFi​j\\small z\_\{ij\}=\\frac\{M\_\{ij\}\-F\_\{ij\}\}\{\\sqrt\{F\_\{ij\}\(1\-p\_\{ij\}\)\}\},\\qquad p\_\{ij\}=\\frac\{F\_\{ij\}\}\{\\sum\_\{j\}F\_\{ij\}\}
Algorithm[7](https://arxiv.org/html/2607.15544#alg7)promptsℳ\\mathcal\{M\}with\(S,P,𝒞∗,𝒟aug\)\(S,P,\\mathcal\{C\}^\{\*\},\\mathcal\{D\}\_\{\\mathrm\{aug\}\}\)requiring a<cot\>block per bullet\. The clean summary𝒞^\\hat\{\\mathcal\{C\}\}is verified viaM1M\_\{1\}\(fidelity\),M2M\_\{2\}\(coverage\), andM3M\_\{3\}\(style similarity\)\.

### B\.1Step 5: Summary Generation and Verification

#### Chain\-of\-thought enforcement\.

The generation prompt \(Algorithm[7](https://arxiv.org/html/2607.15544#alg7)\) requires a<cot\>block before every bullet, with four mandatory fields:ROWS\(comparison IDs cited\),VALUES\(exact pct\-change values from those rows\),CALC\(any arithmetic\), andCHECK\(confirmation that every number in the upcoming bullet appears above\)\. After generation,<cot\>blocks are stripped to produce the clean summary; the raw output with<cot\>blocks is saved separately as an audit log\.

Row\-coverage check\.After generation, a regex pass extracts all comparison IDs cited in<cot\> ROWS:lines and computes coverage =\|cited∩all\_ids\|/\|all\_ids\|\|\\text\{cited\}\\cap\\text\{all\\\_ids\}\|/\|\\text\{all\\\_ids\}\|\. Missing IDs are reported\.

#### M1 \(Factual Fidelity\)\.

An LLM verifier checks every number in the summary against three sources: \(1\) the<cot\>audit log, \(2\) the source comparison table, and \(3\) the GT Key Takeaways\. Flexibility is allowed for rounding \(5%≈4\.8%5\\%\\approx 4\.8\\%\), sign conventions \(−16\.6%\-16\.6\\%reported as “16\.6% reduction”\), and simple derived arithmetic \(e\.g\.,208,683−130,196=78,487208\{,\}683\-130\{,\}196=78\{,\}487\)\.

#### M2 \(GT Coverage\)\.

The evaluator extracts scenario\-based GT findings \(skipping trajectory descriptions and caveats\), labels each as COVERED, PARTIAL, or MISSING\.

#### M3 \(Style Consistency\)\.

Computed deterministically using TF\-IDF cosine similarity \(unigrams \+ bigrams\) and optionally sentence\-embedding cosine similarity \(all\-MiniLM\-L6\-v2\)\. Additional diagnostics include vocabulary Jaccard overlap, average sentence length, and numeric density \(fraction of sentences containing a number\)\.

### B\.2M1 Verification: Detailed Examples

Figure[7](https://arxiv.org/html/2607.15544#A2.F7)in the main text shows twoMISMATCHexamples from the M1 verifier\. We provide additional context:

Example 1: Range exclusion\.The generated summary claims “deaths are reduced by 6\.6% to 8\.4%\.” The source data contains three values:−6\.6%\-6\.6\\%,\+0\.4%\+0\.4\\%,−8\.4%\-8\.4\\%\. The value\+0\.4%\+0\.4\\%represents an*increase*, not a reduction, so the stated range\[6\.6%,8\.4%\]\[6\.6\\%,8\.4\\%\]is misleading—it excludes a data point that reverses the direction\.

Example 2: Incomplete range\.The summary states “hospitalizations are reduced by 6\.3% to 7\.4%\.” The source values are\+6\.2%\+6\.2\\%,\+6\.3%\+6\.3\\%,−7\.4%\-7\.4\\%\. The range\[6\.3%,7\.4%\]\[6\.3\\%,7\.4\\%\]omits6\.2%6\.2\\%and conflates increases with decreases \(the−7\.4%\-7\.4\\%is a reduction; the\+6\.2%\+6\.2\\%and\+6\.3%\+6\.3\\%are increases\)\.

These examples illustrate a systematic failure mode: when derived statistics \(ranges, averages\) are computed over values with mixed signs, the resulting claim can be technically “close” to the data but semantically incorrect\.

![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/M1_issue.png)Figure 7:Example M1 \(factual fidelity\) verification output showing two flagged claims from a generated summary\. In both cases the verifier marks the claim asMISMATCH: the first because the reported range excludes a positive value present in the source data, and the second because the stated range does not capture all three source values\. These examples illustrate how derived statistics \(ranges, averages\) can introduce numerical discrepancies even when the underlying point estimates are correct\.
### B\.3Algorithm

Algorithm 1Step 0: Style Extraction1:Previous\-round GT reports

\{Gr\}r≠r∗\\\{G\_\{r\}\\\}\_\{r\\neq r^\{\*\}\}
2:Style description

SS
3:foreach round

r≠r∗r\\neq r^\{\*\}do

4:

gr←ExtractKeyTakeaways​\(Gr\)g\_\{r\}\\leftarrow\\textsc\{ExtractKeyTakeaways\}\(G\_\{r\}\)
5:endfor

6:

S←ℳ​\(StylePrompt,\{gr\}\)S\\leftarrow\\mathcal\{M\}\\\!\\bigl\(\\textsc\{StylePrompt\},\\;\\\{g\_\{r\}\\\}\\bigr\)

Algorithm 2Step 1a: Axis Decomposition \(Partial Order\)1:README text

RR, flagfallback

2:Decomposition

P=\(𝒜,ρ,\{𝐭s\}s∈𝒮\)P=\(\\mathcal\{A\},\\;\\rho,\\;\\\{\\mathbf\{t\}\_\{s\}\\\}\_\{s\\in\\mathcal\{S\}\}\)⊳\\trianglerightaxes, roles, per\-scenario axis values

3:

𝑟𝑜𝑤𝑠←ParseScenarioTable​\(R\)\\mathit\{rows\}\\leftarrow\\textsc\{ParseScenarioTable\}\(R\)
4:

𝑑𝑒𝑠𝑐←ParseDescriptions​\(R\)\\mathit\{desc\}\\leftarrow\\textsc\{ParseDescriptions\}\(R\)
5:ifnotfallbackthen

6:

P←ℳ​\(DecomposePrompt,𝑟𝑜𝑤𝑠,𝑑𝑒𝑠𝑐\)P\\leftarrow\\mathcal\{M\}\\\!\\bigl\(\\textsc\{DecomposePrompt\},\\;\\mathit\{rows\},\\;\\mathit\{desc\}\\bigr\)
7:if

PPvalid and completethenreturn

PP
8:endif

9:endif

10:

P←FallbackDecompose​\(𝑟𝑜𝑤𝑠\)P\\leftarrow\\textsc\{FallbackDecompose\}\(\\mathit\{rows\}\)⊳\\trianglerightsplit names on ‘\_’, all roles←\\leftarrowprimary

Algorithm 3Step 1b: Agentic Schema Inference1:Parquet file

DrawD\_\{\\mathrm\{raw\}\}
2:Preprocessing state

σ\\sigmawith validated column mapping

3:

σ\(0\)←Draw∪ℳ​\(SchemaPrompt,Describe​\(Draw\)\)\\sigma^\{\(0\)\}\\leftarrow D\_\{\\mathrm\{raw\}\}\\cup\\mathcal\{M\}\\\!\\bigl\(\\textsc\{SchemaPrompt\},\\;\\textsc\{Describe\}\(D\_\{\\mathrm\{raw\}\}\)\\bigr\)
4:for

t=1,…,Tt=1,\\dots,Tdo

5:Apply

σ\(t−1\)\\sigma^\{\(t\-1\)\}to

DrawD\_\{\\mathrm\{raw\}\}: rename, cast, filter, aggregate

6:ifsuccessthenreturn

σ\(t−1\)\\sigma^\{\(t\-1\)\}
7:else

8:

σ\(t\)←ℳ​\(σ\(t−1\),ϵ\(t−1\)\)\\sigma^\{\(t\)\}\\leftarrow\\mathcal\{M\}\\\!\\bigl\(\\sigma^\{\(t\-1\)\},\\;\\epsilon^\{\(t\-1\)\}\\bigr\)⊳\\trianglerightLLM diagnoses tracebackϵ\\epsilon

9:endif

10:endfor

Algorithm 4Step 2: Dataset Augmentation \(Union Table\)1:

PP, preprocessed data

DD
2:Augmented dataset

DaugD\_\{\\mathrm\{aug\}\}
3:

𝑠𝑡𝑎𝑡𝑠←AgentPreprocess​\(D\)\\mathit\{stats\}\\leftarrow\\textsc\{AgentPreprocess\}\(D\)⊳\\trianglerightper \(scenario, age, target\): median

4:for

r=1r=1to

\|𝒮\|\|\\mathcal\{S\}\|do

5:foreach

C∈\(𝒮r\)C\\in\\binom\{\\mathcal\{S\}\}\{r\}do

6:

𝑐𝑜𝑚𝑚𝑜𝑛←⋂s∈C𝐭s\\mathit\{common\}\\leftarrow\\bigcap\_\{s\\in C\}\\mathbf\{t\}\_\{s\}⊳\\trianglerightshared axis values

7:

𝑢𝑛𝑖𝑜𝑛←⋃s∈C𝐭s\\mathit\{union\}\\leftarrow\\bigcup\_\{s\\in C\}\\mathbf\{t\}\_\{s\}
8:foreach age group

aa, target

qqdo

9:Compute

min\\min,

max\\max,

mean\\mathrm\{mean\}of member medians

10:Append row to

DaugD\_\{\\mathrm\{aug\}\}
11:endfor

12:endfor

13:endfor

Algorithm 5Step 3: Fact Generation \(Comparison Grammar\)1:

P=\(𝒜,ρ,\{𝐭s\}\)P=\(\\mathcal\{A\},\\rho,\\\{\\mathbf\{t\}\_\{s\}\\\}\), augmented data

DaugD\_\{\\mathrm\{aug\}\}
2:Candidate fact set

𝒞\\mathcal\{C\}
3:

𝒮cf←\{s:s​is counterfactual\}\\mathcal\{S\}\_\{\\mathrm\{cf\}\}\\leftarrow\\\{s:s\\text\{ is counterfactual\}\\\};

𝒮¬←𝒮∖𝒮cf\\mathcal\{S\}\_\{\\neg\}\\leftarrow\\mathcal\{S\}\\setminus\\mathcal\{S\}\_\{\\mathrm\{cf\}\}
4:

5:// R1: Marginal Pairs

6:for

\(sa,sb\)∈\(𝒮¬2\)\(s\_\{a\},s\_\{b\}\)\\in\\binom\{\\mathcal\{S\}\_\{\\neg\}\}\{2\}do

7:if

\|\{k:𝐭sak≠𝐭sbk\}\|=1\|\\\{k:\\mathbf\{t\}\_\{s\_\{a\}\}^\{k\}\\neq\\mathbf\{t\}\_\{s\_\{b\}\}^\{k\}\\\}\|=1then

8:

𝒞←𝒞∪\{MP​\(sa,sb\)\}\\mathcal\{C\}\\leftarrow\\mathcal\{C\}\\cup\\\{\\textsc\{MP\}\(s\_\{a\},s\_\{b\}\)\\\}
9:endif

10:endfor

11:

12:// R2: Grouped Counterfactual

13:foreach primary axis

kkwith

ρ​\(k\)=primary\\rho\(k\)=\\texttt\{primary\}do

14:foreach value

vv, group

Gv←\{s∈𝒮¬:𝐭sk=v\}G\_\{v\}\\leftarrow\\\{s\\in\\mathcal\{S\}\_\{\\neg\}:\\mathbf\{t\}\_\{s\}^\{k\}=v\\\}do

15:ifwithin\-group variation only on modifier axesthen

16:

𝒞←𝒞∪\{GC​\(𝒮cf,Gv\)\}\\mathcal\{C\}\\leftarrow\\mathcal\{C\}\\cup\\\{\\textsc\{GC\}\(\\mathcal\{S\}\_\{\\mathrm\{cf\}\},\\;G\_\{v\}\)\\\}
17:endif

18:endfor

19:endfor

20:

21:// R3: Second\-Degree \(Pool vs Pool\)

22:foreach axis

kk, each pair of values

\(va,vb\)\(v\_\{a\},v\_\{b\}\)do

23:

Ga←\{s:𝐭sk=va\}G\_\{a\}\\leftarrow\\\{s:\\mathbf\{t\}\_\{s\}^\{k\}=v\_\{a\}\\\},

Gb←\{s:𝐭sk=vb\}G\_\{b\}\\leftarrow\\\{s:\\mathbf\{t\}\_\{s\}^\{k\}=v\_\{b\}\\\}
24:

𝑏𝑎𝑙𝑎𝑛𝑐𝑒𝑑←\\mathit\{balanced\}\\leftarrow\(other\-axis distributions match\)

25:

𝒞←𝒞∪\{SD​\(Ga,Gb,𝑏𝑎𝑙𝑎𝑛𝑐𝑒𝑑\)\}\\mathcal\{C\}\\leftarrow\\mathcal\{C\}\\cup\\\{\\textsc\{SD\}\(G\_\{a\},G\_\{b\},\\mathit\{balanced\}\)\\\}
26:endfor

27:

28:// Attach numerics

29:foreach

c∈𝒞c\\in\\mathcal\{C\}, each age group

aa, target

qqdo

30:

Δca,q←\(vB−vA\)/\|vA\|×100\\Delta\_\{c\}^\{a,q\}\\leftarrow\(v\_\{B\}\-v\_\{A\}\)/\|v\_\{A\}\|\\times 100⊳\\triangleright% increment/decrement fromDaugD\_\{\\mathrm\{aug\}\}

31:endfor

Algorithm 6Step 4: Interestingness Filtering \(MaxEnt / IPF\)1:Observation matrix

M∈ℝN×PM\\in\\mathbb\{R\}^\{N\\times P\}\(states

×\\timesfields\)

2:Selected subset

𝒞∗⊆𝒞\\mathcal\{C\}^\{\*\}\\subseteq\\mathcal\{C\}
3:

M~←ZScore​\(M\)\\tilde\{M\}\\leftarrow\\textsc\{ZScore\}\(M\)⊳\\trianglerightcolumn\-normalise

4:

F←IPF​\(𝟏N×P,Σregion,Σcategory,Σmetric\)F\\leftarrow\\textsc\{IPF\}\\\!\\bigl\(\\mathbf\{1\}\_\{N\\times P\},\\;\\Sigma\_\{\\mathrm\{region\}\},\\;\\Sigma\_\{\\mathrm\{category\}\},\\;\\Sigma\_\{\\mathrm\{metric\}\}\\bigr\)⊳\\trianglerightuniform prior→\\tomarginals

5:foreach cell

\(i,j\)\(i,j\)do

6:

pi​j←Fi​j/∑jFi​jp\_\{ij\}\\leftarrow F\_\{ij\}\\,/\\,\\sum\_\{j\}F\_\{ij\}
7:

zi​j←Mi​j−Fi​jFi​j​\(1−pi​j\)z\_\{ij\}\\leftarrow\\dfrac\{M\_\{ij\}\-F\_\{ij\}\}\{\\sqrt\{F\_\{ij\}\\,\(1\-p\_\{ij\}\)\}\}⊳\\trianglerightmultinomialzz\-score

8:endfor

9:

𝒞∗←BalancedTopK​\(\{\|zi​j\|\},K\)\\mathcal\{C\}^\{\*\}\\leftarrow\\textsc\{BalancedTopK\}\(\\\{\|z\_\{ij\}\|\\\},\\;K\)⊳\\trianglerightcap⌊K/2⌋\\lfloor K/2\\rfloorper field

10:

HK←−∑ip^i​log⁡p^iH\_\{K\}\\leftarrow\-\\sum\_\{i\}\\hat\{p\}\_\{i\}\\log\\hat\{p\}\_\{i\}⊳\\trianglerightselection entropy

Algorithm 7Step 5: Summary Generation & Verification1:

SS,

PP,

𝒞∗\\mathcal\{C\}^\{\*\},

DaugD\_\{\\mathrm\{aug\}\}, scenario descriptions

2:Summary

C^\\hat\{C\}, verification scores

\(M1,M2,M3\)\(M\_\{1\},M\_\{2\},M\_\{3\}\)
3:

4:// 5a: Generate

5:

𝑝𝑟𝑜𝑚𝑝𝑡←\(S,P,𝒞∗,Daug,𝑐ℎ𝑒𝑐𝑘𝑙𝑖𝑠𝑡\)\\mathit\{prompt\}\\leftarrow\(S,\\;P,\\;\\mathcal\{C\}^\{\*\},\\;D\_\{\\mathrm\{aug\}\},\\;\\mathit\{checklist\}\)
6:

C^raw←ℳ​\(𝑝𝑟𝑜𝑚𝑝𝑡\)\\hat\{C\}\_\{\\mathrm\{raw\}\}\\leftarrow\\mathcal\{M\}\(\\mathit\{prompt\}\)⊳\\trianglerightwith mandatory<cot\>per bullet

7:

C^←StripCoT​\(C^raw\)\\hat\{C\}\\leftarrow\\textsc\{StripCoT\}\(\\hat\{C\}\_\{\\mathrm\{raw\}\}\)
8:

9:// 5b: Verify

10:

M1←Nverified/NclaimsM\_\{1\}\\leftarrow N\_\{\\mathrm\{verified\}\}\\;/\\;N\_\{\\mathrm\{claims\}\}⊳\\trianglerightfactual fidelity via LLM

11:

M2←\(Ncovered\+12​Npartial\)/NGTM\_\{2\}\\leftarrow\(N\_\{\\mathrm\{covered\}\}\+\\tfrac\{1\}\{2\}N\_\{\\mathrm\{partial\}\}\)\\;/\\;N\_\{\\mathrm\{GT\}\}⊳\\trianglerightGT coverage

12:

M3←cos⁡\(eC^,eGT\)M\_\{3\}\\leftarrow\\cos\(e\_\{\\hat\{C\}\},\\;e\_\{\\mathrm\{GT\}\}\)⊳\\trianglerightstyle similarity

![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/error_woStructure.png)Figure 8:Iterative summary refinement\. Left: the initial Gemma 3 27B output after one round of generation, showing the raw summary, and the annotations highlight specific issues: missing scenario names, inaccurate claims, and wrong age\-group comparison\. Right: feedback on the summary, however, the verification alone can’t identify the mistakes, hence we need the structured representation and cot\.

## Appendix CCase Study: Round 19 Scenario Design

Round 19 of the COVID\-19 Scenario Modeling Hub evaluates the impact of annual booster vaccination strategies for the 2025–26 season\. Five scenarios \(A–E\) are defined along two axes \(Figure[10](https://arxiv.org/html/2607.15544#A3.F10)\):

- •Booster target\(primary axis\):none\(Scenario A, the no\-vaccination counterfactual\),high\_risk\(adults 65\+ and immunocompromised; Scenarios B, C\), orall\_eligible\(all currently eligible age groups; Scenarios D, E\)\.
- •Immunization timing\(modifier axis\):classic\(vaccination campaign beginning mid\-August, consistent with prior seasons; Scenarios B, D\) orearly\(campaign shifted earlier, starting late June; Scenarios C, E\)\.

This yields a2×22\\times 2factorial over the intervention scenarios plus a shared counterfactual, enabling three types of comparison via the grammar\. The four marginal pairs \(e\.g\. B vs\. C, isolating the effect of timing within the high\-risk target\) each differ on exactly one axis\. Two grouped\-counterfactual comparisons pool over timing \(a modifier axis\) to estimate the overall effect of each target population against the baseline\. One second\-degree comparison contrasts the high\-risk pool \{B, C\} against the all\-eligible pool \{D, E\}, balanced on timing\. The decomposition agent classifiesbooster\_targetas primary andimmunization\_timingas modifier, which activates the R2 guard: pooling over timing is permitted \(it is a modifier\), but pooling over target population is blocked in grouped\-CF comparisons \(it is primary\)\. Figure[5](https://arxiv.org/html/2607.15544#A1.F5)and[6](https://arxiv.org/html/2607.15544#A1.F6)show a partial image of the augmented table and the filtered entry, respectively\.

### C\.1Chain\-of\-Thought Narrative: Role and Utility

A common failure mode of LLM\-generated summaries is*untraceable arithmetic*: the model reports a number that cannot be verified against any source row, either because it was hallucinated or because the derivation path is opaque\. The mandatory<cot\>block addresses this by requiring the model to show its work before every bullet \(Figure[9](https://arxiv.org/html/2607.15544#A3.F9)\)\.

Each<cot\>block contains four fields:

1. 1\.ROWS: the comparison IDs \(e\.g\.GC\_005,MP\_001\) from the filtered table \(Table[6](https://arxiv.org/html/2607.15544#A1.F6)\)𝒞\\mathcal\{C\}that the upcoming bullet draws on\.
2. 2\.VALUES: the exact percentage\-change values copied from those rows \(e\.g\.−16\.6%\-16\.6\\%,−14\.3%\-14\.3\\%\)\.
3. 3\.CALC: any arithmetic applied—ranges, means, differences, or absolute\-count derivations \(e\.g\.208,683−130,196=78,487208\{,\}683\-130\{,\}196=78\{,\}487averted hospitalizations\)\.
4. 4\.CHECK: a self\-verification line confirming that every number in the upcoming bullet appears inVALUESorCALC\.

This structure provides three concrete benefits:

1\. Post\-hoc auditability\.The cot output creates a full provenance chain from augmented data to narrative\. A reviewer, either human or automated, can trace any claim backward: bullet→\\to<cot\>→\\tocomparison ID→\\torow in𝒟aug\\mathcal\{D\}\_\{\\mathrm\{aug\}\}→\\tosource parquet\. The M1 verifier exploits this chain directly, checking numbers against theVALUESandCALCfields before falling back to the raw source tables\.

Reduced hallucination\.Requiring the model to cite specific row IDs*before*writing the bullet constrains generation to claims that are grounded in the filtered table\. In the ablation \(Table[3](https://arxiv.org/html/2607.15544#S6.T3)\), removing CoT drops M1 from 100% to 70% while increasing the claim count from 30 to 46—the model produces more numbers but with weaker grounding, because it is no longer forced to anchor each claim to a specific comparison\.

Coverage enforcement\.The row checklist in the prompt lists every comparison ID that must appear in at least oneROWSfield\. After generation, a regex pass computes the fraction of IDs cited\. Missing IDs are reported, and if coverage is below a threshold, the prompt can be reissued with explicit instructions to address the gaps\. This mechanism ensures that the summary does not silently omit important comparisons\.

### C\.2Additional MaxEnt\-based Interestingness Analysis

To identify state\-level patterns that are not apparent from aggregate national summaries alone, we apply a maximum\-entropy interestingness analysis to the state\-level projection outputs\. The MaxEnt baseline is fitted using iterative proportional fitting \(IPF\), simultaneously matching HHS regional marginals \(row axis\) and field\-category/metric marginals \(column axis\)\. All absolute burden and averted\-count fields are expressed per 100,000 population using 2022 ACS five\-year estimates before the matrix is constructed, so that state size does not mechanically inflate residual scores\. Large absolute residuals indicate state–metric cells that are surprising relative to what would be expected from the marginal structure alone\.

1\. State\-level deviation\.After population normalization, the most surprising states overall are concentrated inHHS Region 1 \(New England\): Vermont \(z¯=3\.48\\bar\{z\}=3\.48,max=11\.25\\max=11\.25\), Massachusetts \(z¯=3\.29\\bar\{z\}=3\.29\), New Hampshire \(z¯=3\.23\\bar\{z\}=3\.23\), Maine \(z¯=3\.20\\bar\{z\}=3\.20\), and Rhode Island \(z¯=3\.07\\bar\{z\}=3\.07\)\. Within this cluster, Vermont, New Hampshire, and Maine are primarily driven by higher\-than\-expected rates of averted hospitalizations under Scenario B \(classic timing, high\-risk booster\) relative to their regional baseline, with Vermont reaching74\.274\.2averted hospitalizations per 100,000 population—well above the MaxEnt expectation\. Vermont and New Hampshire additionally show a higher\-than\-expected incremental gain when expanding from Scenario B to Scenario D \(all\-eligible booster, same classic timing\), at30\.930\.9and22\.922\.9per 100,000 respectively, suggesting that broadening booster eligibility yields disproportionate returns in these states\.

West Virginia \(z¯=3\.04\\bar\{z\}=3\.04,max=11\.63\\max=11\.63\) stands out within Region 3 \(Mid\-Atlantic\) as a high\-burden outlier, with a projected median hospitalization rate of281\.2281\.2per 100,000 under Scenario B, significantly exceeding regional expectations\. Kentucky \(220\.5220\.5/100k\) and New York \(246\.8246\.8/100k\) also show higher\-than\-expected hospitalization burden rates\. Conversely, the District of Columbia \(5\.45\.4/100k\), Rhode Island \(10\.910\.9/100k\), and Utah \(5\.95\.9/100k\) exhibit projected death rates that are notably*below*the MaxEnt expectation, representing a distinct cluster of lower\-than\-anticipated mortality burden\.

*Evaluation and justification\.*The selected findings reflect deviations that cannot be explained by regional or metric\-level marginals alone\. Following per\-capita normalization, findings are no longer driven by state population size; instead, they reflect genuine rate\-based heterogeneity across intervention efficacy, hospitalization burden, and mortality concentration\. The New England booster\-responsiveness cluster and the West Virginia high\-burden outlier represent the two primary dimensions of surprise: intervention\-related deviations and absolute burden deviations\. The lower\-than\-expected death rates in DC, Rhode Island, and Utah constitute a third dimension warranting further investigation\.

2\. Age\-specific mortality concentration\.Alaska \(86%\), New Mexico \(85%\), and Mississippi \(87%\) exhibit higher\-than\-expected shares of projected deaths occurring in the 65\+\+age group, all above the national historical baseline of approximately 75–80% for COVID\-19 mortality, despite the 65\+\+population representing only∼\\sim17% of the US population\. These findings are based on relative share measures and are independent of state population size\. Rhode Island shows a complementary signal: a lower\-than\-expected share of*hospitalizations*in the 0–64 age group \(38%\), consistent with a more elderly\-concentrated hospitalization burden\. Complementary age\-share fields \(e\.g\., high 65\+\+share and low 0–64 share for the same state and metric\) are deduplicated in the selection procedure so that each represents a distinct finding\.

3\. MaxEnt selection diagnostics\.To assess whether the selected findings are overly concentrated in one type of signal, we examine the category composition of the top\-15 MaxEnt cells and the per\-field cap applied during selection \(at most⌊N/4⌋=3\\lfloor N/4\\rfloor=3cells per field forN=15N=15\)\. The selected cells span four categories: age\-share \(\|z\|¯=5\.00\\bar\{\|z\|\}=5\.00, 4 cells\), burden \(\|z\|¯=4\.58\\bar\{\|z\|\}=4\.58, 4 cells\), absolute averted \(\|z\|¯=2\.13\\bar\{\|z\|\}=2\.13, 3 cells\), and incremental averted \(\|z\|¯=1\.34\\bar\{\|z\|\}=1\.34, 2 cells\)\. This distribution reflects the category\-level surprise scores: age\-share and burden quantities show the largest mean deviations from the MaxEnt baseline, while percentage\-reduction quantities \(\|z\|¯=2\.56\\bar\{\|z\|\}=2\.56\) and incremental effects are comparatively more consistent with the fitted marginals\. At the regional level, Region 1 \(New England\) exhibits the highest mean surprise \(3\.193\.19\), followed by Region 5 \(Chicago,2\.912\.91\) and Region 3 \(Mid\-Atlantic,2\.902\.90\), pointing to a broader clustering of surprising patterns across the Northeast and Midwest\.

Overall, the MaxEnt analysis provides a complementary view of the ensemble projections by identifying where state\-level patterns deviate most strongly from the expected marginal structure\.

Rather than replacing the national summary, this analysis helps prioritize which state–metric combinations may deserve closer inspection, particularly in the context of resource allocation and targeted public health intervention\.

We further evaluate the selected statements with an epidemiology expert, who scores each statement based on its narrative interestingness\. This evaluation addresses the question:*Are the top\-KKMaxEnt\-selected factors actually interesting to domain experts?*We find that the interestingness\-based selection adds value to the final narrative, since experts prioritize state\-level and regional burden patterns as well as scenario\-wise comparative facts that may otherwise be missed in aggregate summaries\. Figures[12](https://arxiv.org/html/2607.15544#A3.F12),[13](https://arxiv.org/html/2607.15544#A3.F13), and[14](https://arxiv.org/html/2607.15544#A3.F14)show expert ratings for the top 15 MaxEnt\-selected facts\. Across these facts, epidemiology experts rated 33% as interesting enough to retain in the final narrative\.

![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/cot_final.png)Figure 9:Chain\-of\-thought audit trail for the finalEpiNarratesummary \(Round 19\)\. Left column: the clean summary bullets as they appear in the output\. Right column: the corresponding<cot\>blocks showing the comparison IDs cited \(ROWS\), the exact percentage\-change values pulled from the filtered table \(VALUES\), any arithmetic performed \(CALC\), and the consistency check \(CHECK\)\. Highlighted spans show where filtered\-table row IDs map to specific claims in the final text, demonstrating end\-to\-end traceability from augmented data to narrative\.![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/sc19.png)Figure 10:Round 19 Scenarios![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/metrics.png)Figure 11:Round 19 evaluation onEpiNarrate![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/maxent1.png)Figure 12:Expert evaluation of the MaxEnt results \(1\)![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/maxent2.png)Figure 13:Expert evaluation of the MaxEnt results \(2\)![Refer to caption](https://arxiv.org/html/2607.15544v1/Figures/appendix/maxent3.png)Figure 14:Expert evaluation of the MaxEnt results \(3\)