Skip to content

EpiNarrate: Agentic Generation of Grounded Narratives from Epidemiological Scenario Projections

🕒 Published (v1): 2026-07-17 01:30 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

EpiNarrate is a five-stage agentic pipeline that generates quantitatively grounded public-health narratives from large multidimensional epidemiological ensemble datasets by separating structured numerical reasoning from natural-language generation. It induces a partial-order (poset) schema over scenario axes, enforces a comparison grammar to prevent invalid contrasts, and selects high-salience facts via maximum-entropy (MaxEnt/IPF) filtering before handing a compact, verified fact set to an LLM for prose generation. On COVID-19 Scenario Modeling Hub data it achieves 100% factual fidelity (M1) on up to 34 numerical claims per round while outperforming all baselines on coverage (M2).

Problem

Direct LLM summarization of large epidemiological ensemble datasets—structured as K-way tensors with \(\sim\!1.6\times10^6\) entries across scenarios, outcomes, subgroups, regions, weeks, and quantiles—produces hallucinated statistics, invalid cross-axis comparisons, and uneven coverage. Standard D2T and prompting approaches (LIDA, single-shot CoT, plan-then-generate) lack structural constraints that enforce arithmetic correctness or prevent "apples-to-oranges" scenario contrasts.

Method

A five-stage modular pipeline, preceded by a style-extraction step (Step 0):

  1. Axis decomposition (Step 1): A decomposition agent \(M_D\) parses free-text scenario descriptions into structured tuples \(t \in V_1\times\cdots\times V_K\), distinguishing primary from modifier axes and constructing a product poset \(\leq_i\) over each attribute. Schema loading is self-correcting: on validation error \(\epsilon^{(t)}\), the schema is iteratively refined as \(\sigma^{(t+1)} \leftarrow M(\sigma^{(t)}, \epsilon^{(t)})\) for up to \(T\) rounds.

  2. Dataset augmentation (Step 2): Deterministic relational operations (e.g., union over poset-parent nodes) extend the base dataset \(D\) to \(D_\text{aug}\), materialising aggregate statistics that do not appear explicitly in the source tables.

  3. Fact generation via comparison grammar (Step 3): A grammar restricts valid comparisons to structurally sound pairs—e.g., rule R1: scenarios differ on exactly one axis; R2: baseline vs. a pool sharing a common secondary axis value. Each fact is associated with a deterministic executable computation path \(\pi_i\) over \(D\).

  4. MaxEnt interestingness filtering (Step 4): Each comparison \(c\) carries a feature profile \(\phi(c)\). Iterative Proportional Fitting (IPF) fits a maximum-entropy distribution \(F\) over a state-by-field matrix \(M \in \mathbb{R}^{N\times P}\) (column-normalised: \(\tilde{M}_{ij}=(M_{ij}-\mu_j)/\sigma_j\)), yielding residuals \(R_{ij}=\tilde{M}_{ij}-F_{ij}\) that identify surprising cells. Selection entropy \(H_K = -\sum_i \hat{p}_i\log\hat{p}_i\) penalises geographic redundancy.

  5. Narrative generation (Step 5): A three-step CoT chain (Observe → Reflect → Synthesize) converts the top-\(K\) verified facts into prose, keeping pattern detection, interpretation, and communication in separate logged steps.

Key Contributions

  • Formal problem statement for public-health narrative generation requiring simultaneous correctness, coverage, and interestingness/non-redundancy.
  • Partial-order-guided schema induction from free-text scenario descriptions with iterative self-repair, enabling robust handling of inconsistent data formats.
  • Comparison grammar that enforces semantically and arithmetically valid contrasts over multidimensional ensemble projections, replacing unconstrained LLM generation.
  • MaxEnt/IPF-based interestingness selection that discourages redundancy while surfacing underrepresented cells, with human expert validation.
  • Three complementary evaluation metrics: M1 (factual fidelity via LLM-based claim verification), M2 (GT key-takeaway coverage), M3 (embedding cosine style similarity).

Results

All scores from Table 1 on COVID-19 SMH rounds R17–R19:

  • EpiNarrate M1: 100%/30 (R19), 89%/34 (R18), 100%/26 (R17) — highest claim count and among highest accuracy of all systems.
  • GPT 5.5 M1: 100%/13 (R19), 88%/28 (R18), 100%/15 (R17) — comparable accuracy but roughly half the claim count.
  • EpiNarrate M2 (coverage): 0.78 (R19), 0.44 (R18), 0.62 (R17) — highest across all rounds vs. all baselines.
  • Single-shot CoT M1: 0%/4 (R19), 12%/8 (R18), 83%/6 (R17); M2: 0.00 across R19 and R18.
  • LIDA M1: 38%/8 (R19), 20%/10 (R18), 25%/8 (R17); M2: 0.00 on R18 and R17.
  • EpiNarrate M3 (style): 0.69–0.81, comparable to GPT 5.5 (0.79–0.86), well above encoder-decoder baselines (0.06–0.48).
  • Full pipeline runtime: 3–5 minutes per round (API-latency dominated), no fine-tuning required.

Limitations

  • M1 is LLM-evaluated and thus stochastic; a correctly derived value can be marked incorrect if the verifier fails to trace arithmetic, making M1 an approximate rather than exact measure.
  • A single correct claim yields a perfect M1 score, so M1 does not capture quantitative richness—must be read alongside claim count.
  • Ground-truth (GT) summaries are not exhaustive gold standards; MaxEnt-surfaced insights absent from GT are penalised by lexical metrics (BLEU/ROUGE), creating an evaluation blind spot.
  • All LLM calls use Gemma 4 26B/31B at temperature 0.0–0.4 via a free-tier API; latency and quota constraints limit throughput.
  • Evaluation covers only COVID-19 SMH rounds; generalisability to other pathogens or non-SMH ensemble formats is unverified.
  • The paper text provided is truncated; MaxEnt human-rating details and additional ablations are deferred to appendices not fully available here.

Relevance to Harnesses / Meta-Harnesses

EpiNarrate is a concrete, domain-specific example of a multi-agent harness in which several specialized agents (decomposition agent, comparison-selection agent, augmentation agent, CoT summarization chain) are orchestrated through a fixed, deterministic pipeline with structured intermediate representations acting as guardrails between stages. The iterative self-repair loop for schema inference (\(\sigma^{(t+1)} \leftarrow M(\sigma^{(t)},\epsilon^{(t)})\)) and the three-step Observe→Reflect→Synthesize chain are direct harness-level control mechanisms that replace unconstrained end-to-end LLM calls. For researchers tracking harnesses, this paper demonstrates how inserting grammar-enforced intermediate structures and entropy-based selection between agent stages can dramatically reduce hallucination and improve coverage without any fine-tuning—a design principle transferable to harnesses in other high-stakes, data-intensive domains.