Data2Story: Virtual Newsroom Agent

• Data2Story is an agentic system that transforms raw data into complete, verifiable news stories through iterative search, planning, and narrative integration.
• The framework employs specialized roles (Detective, Analyst, Editor, Designer, Programmer, Auditor, Inspector) to coordinate evidence retrieval, analysis, and multimodal asset generation.
• Evaluations highlight enhanced transparency and precision with metrics demonstrating improved claim traceability and consistent journalistic quality.

Data Journalist Agent, often abbreviated Data2Story, denotes an agentic system that transforms raw data, retrieved context, and multimodal evidence into a complete, verifiable news story. In the literature, the term refers both to an implementation-oriented blueprint derived from journalism benchmarks such as NEWSAGENT and to a 2026 multi-agent framework that organizes a virtual newsroom around specialized roles including Detective, Analyst, Editor, Designer, Programmer, Auditor, and Inspector (Chien et al., 30 Aug 2025, Lin et al., 9 Jun 2026). Its defining departure from one-shot summarization is the requirement to iteratively search, plan, analyze, edit, verify, and present a narrative whose claims remain traceable to data, code, or external references.

1. Research setting and conceptual scope

Data2Story emerges from a convergence of work on agentic newswriting, data-driven storytelling, investigative reporting support, and provenance-aware verification. NEWSAGENT frames the core problem as whether modern agent frameworks can act as journalists under real newsroom constraints: agents begin with a title, simulated release date, and firsthand objects, then must identify narrative perspectives, issue keyword-based queries, retrieve historical background available before publication, edit a draft, and rephrase it into a complete article (Chien et al., 30 Aug 2025). This formulation explicitly models journalism as an iterative, exploratory workflow rather than a static summarization task.

The 2026 Data Journalist Agent framework extends that agenda from text-first news drafting to end-to-end multimedia reporting. Its central claim is that a newsroom-grade agent must satisfy two conditions simultaneously: claims must be evidence-grounded, and articles must be multimodally generative, selecting maps, audio, video, charts, or other assets according to what readers need to inspect (Lin et al., 9 Jun 2026). In that sense, Data2Story is not merely a text generator with retrieval; it is a coordinated production system whose outputs include executable analyses, editorial structure, visual assets, interaction logic, and provenance bindings.

A plausible implication is that Data2Story is best understood as a virtual newsroom rather than a single model. That interpretation is supported by both lines of work: NEWSAGENT emphasizes perception–action loops, time-aware search, and structured editing functions (Chien et al., 30 Aug 2025), while Data2Story formalizes distinct newsroom roles with explicit artifact handoffs (Lin et al., 9 Jun 2026).

2. Virtual newsroom architecture

The 2026 instantiation of Data2Story defines seven specialized roles. Detective augments the raw dataset $D$ with web-found context $\hat{Z}$; Analyst writes executable code $C$ and produces results $R$; Editor selects angles and produces a paragraph-level outline $F$; Designer generates multimodal assets $V$; Programmer assembles the final HTML/CSS/JS article $U$; Auditor reviews the live page for defects; and Inspector decomposes the final article into fragments and binds each fragment to upstream evidence $E = D \cup R \cup C \cup F \cup V$ (Lin et al., 9 Jun 2026). The Programmer is constrained to assembly and revision rather than factual invention, which enforces separation between analysis and presentation.

NEWSAGENT provides a complementary module view that is implementation-oriented rather than role-oriented. Its proposed Data2Story blueprint includes task decomposition and planning, query expansion, multi-hop time-aware retrieval, timeline construction, entity resolution and coreference, evidence aggregation with citation tracking, drafting and iterative editing, style and tone control, fact-checking and contradiction detection, and an ethics/safety module (Chien et al., 30 Aug 2025). Together, these two descriptions define the same system at different granularities: one as human-like newsroom roles, the other as reusable agentic subsystems.

Role or module	Core function	Primary artifact
Detective / retrieval stack	Find contextual evidence and background	$\hat{Z}$
Analyst / analysis stack	Execute code over data and context	$C$, $\hat{Z}$0
Editor / planning stack	Choose angle and structure	$\hat{Z}$1
Designer / multimodal stack	Create charts, maps, audio, video, interactives	$\hat{Z}$2
Programmer + Auditor	Assemble and repair the live article	$\hat{Z}$3, $\hat{Z}$4
Inspector / provenance stack	Bind article fragments to evidence	claim–evidence map

The architecture is explicitly provenance-rich. In the 2026 framework, every result $\hat{Z}$5 carries a pointer to the exact script that generated it, each editorial finding is annotated with upstream results and code, and each generated asset stores the full tool call and parameters (Lin et al., 9 Jun 2026). NEWSAGENT’s extension toward Data2Story adds provenance metadata such as source, time, and confidence inside agent memory, alongside a task-level journal of actions and rationale (Chien et al., 30 Aug 2025).

3. Iterative workflow, data model, and retrieval

NEWSAGENT supplies the most concrete task model for Data2Story-like systems. The benchmark contains approximately 6.3k human-verified tasks derived from BBC and APNews articles, with two validated counts reported during curation, 6,237 and 6,327 (Chien et al., 30 Aug 2025). Each task is built from text objects under a unified JSON schema: descriptions from article body text, captions prefixed with "[Caption]", and transcript turns prefixed with speaker names. Objects are split into firsthand and historical material, with a reported distribution of 69% firsthand and 31% historical (Chien et al., 30 Aug 2025).

The task interface exposes only the title, release date, and firsthand objects. Agents must then operate through a constrained action space: Search, Insert, Remove, Terminate, and Rephrase. Search queries a historical database strictly before the release date and returns top-$\hat{Z}$6 results with $\hat{Z}$7 and cosine similarity greater than $\hat{Z}$8 using all-MiniLM-L6-v2 embeddings; Insert can add only previously retrieved objects; Remove deletes objects already in the draft; invalid actions count against a global budget of at most 20 operations per task (Chien et al., 30 Aug 2025). Two execution modes are evaluated: a 1-step mode that specifies operation and parameters in one turn, and a 2-step mode that first chooses the operation and then specifies parameters.

This action model is important because it operationalizes journalism as controlled evidence manipulation. After Terminate, the system rewrites textual components of the draft into a coherent article while preserving links to images and transcripts through object references (Chien et al., 30 Aug 2025). The proposed Data2Story algorithm in the same synthesis makes this more explicit: initialize a draft $\hat{Z}$9 from core firsthand objects, plan an outline $C$0, maintain memory $C$1, analyze gaps $C$2, then iteratively Search, Insert, Remove, self-reflect, and terminate once outline coverage is sufficient (Chien et al., 30 Aug 2025). This suggests that Data2Story’s distinctive competence lies not only in retrieval but in narrative integration under uncertainty.

The 2026 Data2Story paper preserves this iterative logic while broadening the input space beyond converted text. Detective first searches the web for context, Analyst executes code over $C$3, Editor ranks findings, Designer chooses modalities, Programmer assembles the story, Auditor repairs defects, and Inspector binds the final article to evidence (Lin et al., 9 Jun 2026). Where NEWSAGENT abstracts newsroom functions into symbolic actions, Data2Story instantiates them as specialized agents and typed artifacts.

4. Verifiability and evaluation

Evaluation in Data2Story research operates at two levels: function-wise task performance and end-to-end journalistic quality. NEWSAGENT measures Search and Edit against journalist-selected ground-truth objects using standard precision, recall, and $C$4:

$C$5

It also reports retrieval-oriented measures such as Precision@$C$6, Recall@$C$7, MRR, and nDCG@$C$8 (Chien et al., 30 Aug 2025).

For end-to-end evaluation, NEWSAGENT uses pairwise GPT-4 comparative judgment on six journalistic dimensions: Factual Consistency, Logical Consistency, Importance, Readability, Objectivity, and Journalistic Style. The paper reports that this pairwise protocol achieved 72% agreement with human judgments, compared with 53% for a single-turn GPT-4 baseline (Chien et al., 30 Aug 2025). The Data2Story extension adds generation metrics such as ROUGE-N, ROUGE-L, BLEU, and BERTScore, along with factuality-oriented measures including knowledge-grounding checks, citation precision/recall, contradiction detection, temporal consistency, outline coherence score, claim verification rate, and editor pass rate (Chien et al., 30 Aug 2025).

The 2026 Data2Story evaluation introduces a broader four-axis protocol over 18 article pairs. Human–agent angle coverage is measured as

$C$9

and the reported headline results are 50.4% of human claims covered by the agent and 35.1% of agent claims appearing in the human article (Lin et al., 9 Jun 2026). A rubric study with 53 participants scores five dimensions—Visual design, Narrative & pacing, Data & method transparency, Claim–data alignment, and Insight value—and reports an overall mean of 4.21 for Data2Story versus 3.38 for human references, with overall preference counts of 39 for Data2Story, 13 for human articles, and 1 tie (Lin et al., 9 Jun 2026).

Verifiability is treated as a separate property. For computational claims, the verifier recomputes $R$0 and accepts the claim if $R$1; for reference-supported claims, it checks support against the cited URL. The article-level verifiability rate is

$R$2

(Lin et al., 9 Jun 2026). On the auditability dimension, the paper reports that 93% of Data2Story claims have machine-checkable provenance bindings, compared with 25% for human-written references (Lin et al., 9 Jun 2026). This is the clearest formalization of what “verifiable journalism agent” means in this literature: not only factually plausible output, but claim-level re-executability or reference traceability.

5. Empirical findings and methodological lineage

The NEWSAGENT results show that current agents are better at retrieving relevant facts than at integrating them into a coherent narrative. Remove operations were never invoked across models, indicating limited self-correction; 2-step mode increased Search counts but did not increase Insert counts; and function-wise scores remained low, with representative 1-step Search $R$3 values of 0.233 for GPT-4o, 0.231 for GPT-4o-mini, 0.206 for Gemma-3-27b-it, and 0.058 for the rule-based baseline (Chien et al., 30 Aug 2025). In 2-step mode, precision rose sharply but recall collapsed, as in Qwen3-32B with precision 0.844, recall 0.071, and $R$4 0.120 (Chien et al., 30 Aug 2025). The paper interprets this as a planning and narrative integration problem rather than a pure retrieval failure.

The 2026 Data2Story results are more favorable at the article level but preserve the same asymmetry. Human articles retain an edge in editorial angle, creative design, and presentation, while Data2Story shows particular strength in transparency and auditability (Lin et al., 9 Jun 2026). The Inspector contributes most strongly to perceived transparency: under a computer-use agent judge, the overall mean rises from 4.60 to 5.10 when the Inspector is available, with Transparency increasing from 4.28 to 5.94 (Lin et al., 9 Jun 2026).

Data2Story also sits within a broader methodological lineage. ConnectionLens demonstrates graph integration of structured, semistructured, and unstructured data into a provenance-rich graph for data journalism, with support for CSV, JSON, XML, HTML, RDF, PDF, NER/NED, and keyword search over answer trees (Anadiotis et al., 2020). DMINR contributes a provenance-first design for verification and exploration through multi-source search, entity extraction, and force-directed connection graphs (MacFarlane et al., 2022). DataScout adds stance-aware retrieval trees that support and oppose a statement through query decomposition, text-to-SQL, and Chain-of-Thought fact extraction (Chen et al., 24 Apr 2025). Compendia organizes quantitative facts extracted from online article collections into thematic clusters and merged Narrative Units for scrollytelling (Karunathilaka et al., 7 Feb 2026), while DataWeaver integrates visualization-to-text and text-to-visualization authoring through call-out interactions and fact anchoring (Fu et al., 29 Mar 2025). TeleFlash contributes a practical LLM-driven reporting system for filtering, summarizing, translating, and distributing daily reports from 170+ Telegram channels, with message-level citations and Slack delivery (Maltezos et al., 25 Aug 2025). A plausible interpretation is that Data2Story functions as an umbrella synthesis of these strands: provenance graphs, investigative verification, stance-aware retrieval, structured fact organization, integrated visual-text authoring, and workflow automation.

6. Limitations, deployment, and future directions

Several limitations recur across the literature. NEWSAGENT itself is text-only after multimodal conversion, uses a single-agent ReAct setup, lacks explicit citation-integrity scoring, and shows an underused Remove function because agents receive weak error signals for pruning or revising draft structure (Chien et al., 30 Aug 2025). The paper explicitly proposes native multimodality and multi-agent collaboration, including AutoGen- and Tree-of-Thought-style role specialization, as future directions (Chien et al., 30 Aug 2025). The 2026 Data2Story paper likewise reports that human writers remain stronger on reported angles, bespoke design, and dense editorial presentation (Lin et al., 9 Jun 2026).

Deployment introduces additional trade-offs. NEWSAGENT notes that 2-step execution reduces mechanical errors but increases turns, so practical systems should cache embeddings and search results and preserve a bounded operation budget, mirroring the benchmark’s limit of 20 operations per task (Chien et al., 30 Aug 2025). TeleFlash shows that adoption is also an interface problem: despite clear utility for filtering and summarization, delivery through Slack limited uptake, and verification remained the most important unmet need (Maltezos et al., 25 Aug 2025). DMINR and co-design work on digital journalism tools further suggest that newsroom adoption depends on speed, explainability, provenance, and low-friction interfaces rather than raw model capability alone (MacFarlane et al., 2022, Ifrim et al., 2017).

Future development therefore tends to converge on four directions. The first is stronger planning, including outline coverage, angle exploration, and explicit self-reflection for revision (Chien et al., 30 Aug 2025). The second is richer multimodality, reintroducing native images, audio, video, maps, and interaction rather than only text-converted semantics (Chien et al., 30 Aug 2025, Lin et al., 9 Jun 2026). The third is verification infrastructure, including claim-to-evidence binding, independent recomputation, contradiction detection, and editorial audit panels (Lin et al., 9 Jun 2026, MacFarlane et al., 2022). The fourth is newsroom fit: modular connectors, multilingual retrieval, platform-aware outputs, and interfaces that respect deadline-driven work (Maltezos et al., 25 Aug 2025, Ifrim et al., 2017).

In this sense, Data Journalist Agent is less a single settled architecture than a research program. Its stable core is already clear: iterative search over incomplete evidence, code-executed analysis, editorial planning, multimodal presentation, and claim-level provenance. What remains unsettled is how far these components can be integrated without sacrificing the editorial judgment, contextual reporting, and creative presentation that still distinguish expert human journalism (Chien et al., 30 Aug 2025, Lin et al., 9 Jun 2026).